## File: README.md # Agent Skills for Context Engineering A comprehensive, open collection of Agent Skills focused on context engineering and harness engineering principles for building production-grade AI agent systems. These skills teach the art and science of curating context, designing agent operating loops, and evaluating agent behavior across any agent platform. [](https://deepwiki.com/muratcankoylan/Agent-Skills-for-Context-Engineering) ## What is Context Engineering? Context engineering is the discipline of managing the language model's context window. Unlike prompt engineering, which focuses on crafting effective instructions, context engineering addresses the holistic curation of all information that enters the model's limited attention budget: system prompts, tool definitions, retrieved documents, message history, and tool outputs. The fundamental challenge is that context windows are constrained not by raw token capacity but by attention mechanics. As context length increases, models exhibit predictable degradation patterns: the "lost-in-the-middle" phenomenon, U-shaped attention curves, and attention scarcity. Effective context engineering means finding the smallest possible set of high-signal tokens that maximize the likelihood of desired outcomes. ## Recognition This repository is cited in academic research as foundational work on static skill architecture: > "While static skills are well-recognized [Anthropic, 2025b; Muratcan Koylan, 2025], MCE is among the first to dynamically evolve them, bridging manual skill engineering and autonomous self-improvement." 1. [Meta Context Engineering via Agentic Skill Evolution](https://arxiv.org/pdf/2601.21557), Peking University State Key Laboratory of General Artificial Intelligence (2025) 2. [Agent Harness Engineering: A Survey](https://openreview.net/pdf/f358711a95aaaf61fdeffd4ef3fc60fba9b8da57.pdf), CMU, Yale, JHU, NEU, Tulane, UAB, OSU, Virginia Tech, and Amazon (2026) ## Skills Overview ### Foundational Skills These skills establish the foundational understanding required for all subsequent context engineering work. | Skill | Description | |-------|-------------| | [context-fundamentals](skills/context-fundamentals/) | Understand what context is, why it matters, and the anatomy of context in agent systems | | [context-degradation](skills/context-degradation/) | Recognize patterns of context failure: lost-in-middle, poisoning, distraction, and clash | | [context-compression](skills/context-compression/) | Design and evaluate compression strategies for long-running sessions | ### Architectural Skills These skills cover the patterns and structures for building effective agent systems. | Skill | Description | |-------|-------------| | [multi-agent-patterns](skills/multi-agent-patterns/) | Master orchestrator, peer-to-peer, and hierarchical multi-agent architectures | | [long-horizon-prompting](skills/long-horizon-prompting/) | **NEW** Write pseudo-formal task briefs for long-running autonomous agents and parallel orchestrations: exact success predicates, non-counting outcomes, audit-gated return conditions, effort floors, and diversity policies, modeled on the published GPT-5.6 Sol Ultra Cycle Double Cover prompt | | [memory-systems](skills/memory-systems/) | Design short-term, long-term, and graph-based memory architectures | | [tool-design](skills/tool-design/) | Build tools that agents can use effectively | | [filesystem-context](skills/filesystem-context/) | Use filesystems for dynamic context discovery, tool output offloading, and plan persistence | | [hosted-agents](skills/hosted-agents/) | **NEW** Build background coding agents with sandboxed VMs, pre-built images, multiplayer support, and multi-client interfaces | ### Operational Skills These skills address the ongoing operation and optimization of agent systems. | Skill | Description | |-------|-------------| | [context-optimization](skills/context-optimization/) | Apply compaction, masking, and caching strategies | | [latent-briefing](skills/latent-briefing/) | Share task-relevant orchestrator state with workers via task-guided KV cache compaction when the worker runtime is controllable | | [evaluation](skills/evaluation/) | Build evaluation frameworks for agent systems | | [advanced-evaluation](skills/advanced-evaluation/) | Master LLM-as-a-Judge techniques: direct scoring, pairwise comparison, rubric generation, and bias mitigation | | [harness-engineering](skills/harness-engineering/) | Design autonomous agent harnesses with locked metrics, durable logs, novelty gates, rollback, and human approval boundaries | | [self-improvement-loops](skills/self-improvement-loops/) | **NEW** Build loops where the harness itself is the optimization target: RSI, meta-harness search, failure-driven self-edits, evolutionary scaffold search, and acceptance gates for self-modifying systems | ### Development Methodology These skills cover the meta-level practices for building LLM-powered projects. | Skill | Description | |-------|-------------| | [project-development](skills/project-development/) | Design and build LLM projects from ideation through deployment, including task-model fit analysis, pipeline architecture, and structured output design | ### Cognitive Architecture Skills These skills cover formal cognitive modeling for rational agent systems. | Skill | Description | |-------|-------------| | [bdi-mental-states](skills/bdi-mental-states/) | **NEW** Transform external RDF context into agent mental states (beliefs, desires, intentions) using formal BDI ontology patterns for deliberative reasoning and explainability | ## Design Philosophy ### Progressive Disclosure Each skill is structured for efficient context use. At startup, agents load only skill names and descriptions. Full content loads only when a skill is activated for relevant tasks. ### Platform Agnosticism These skills focus on transferable principles rather than vendor-specific implementations. The patterns work across Claude Code, Cursor, and any agent platform that supports skills or allows custom instructions. ### Conceptual Foundation with Practical Examples Scripts and examples demonstrate concepts using Python pseudocode that works across environments without requiring specific dependency installations. ## Usage ### Usage with Claude Code This repository is a **Claude Code Plugin Marketplace** containing context engineering skills that Claude automatically discovers and activates based on your task context. ### Installation **Step 1: Add the Marketplace** Run this command in Claude Code to register this repository as a plugin source: ``` /plugin marketplace add muratcankoylan/Agent-Skills-for-Context-Engineering ``` **Step 2: Install the Plugin** Option A - Browse and install: 1. Select `Browse and install plugins` 2. Select `context-engineering-marketplace` 3. Select `context-engineering` 4. Select `Install now` Option B - Direct install via command: ``` /plugin install context-engineering@context-engineering-marketplace ``` This installs all published skills in a single plugin. Skills are activated automatically based on your task context. ### Skill Activation Scenarios | Skill | Activate When | |-------|---------------| | `context-fundamentals` | Establishing context-window mental models, planning agent architecture, or explaining how context components affect model behavior | | `context-degradation` | Diagnosing attention failures, context poisoning, lost-in-middle behavior, or degraded agent performance across long sessions | | `context-compression` | Preserving useful state while reducing conversation, tool-output, or trajectory size under context pressure | | `context-optimization` | Improving token efficiency, retrieval precision, prefix reuse, masking, partitioning, or budget allocation for agent systems | | `latent-briefing` | Sharing orchestrator trajectory with workers via task-guided KV cache compaction when the worker runtime is controllable and the models are compatible | | `multi-agent-patterns` | Choosing coordination patterns, isolating context across agents, designing handoffs, or evaluating whether parallel agents are justified | | `long-horizon-prompting` | Writing or evaluating the launch prompt for a long-running autonomous agent or parallel orchestration: success predicates, non-counting outcomes, persistence and stop rules, adversarial audit gates, and portfolio diversity policies | | `memory-systems` | Persisting cross-session knowledge, tracking entities over time, choosing memory frameworks, or designing retrieval and update semantics | | `tool-design` | Defining agent-tool contracts, consolidating tool surfaces, improving descriptions, or making tool errors actionable | | `filesystem-context` | Moving large or durable context into files, creating scratchpads, supporting just-in-time discovery, or coordinating agents through shared artifacts | | `hosted-agents` | Running coding agents in remote sandboxes, background environments, warm pools, or multiplayer agent infrastructure | | `evaluation` | Creating deterministic checks, rubrics, regression suites, production monitoring, or quality gates for agent behavior | | `advanced-evaluation` | Using LLM judges, pairwise comparison, calibration, bias mitigation, or human-aligned quality assessment | | `harness-engineering` | Designing autonomous loops with locked evaluators, editable surfaces, durable logs, novelty gates, rollback, and approval boundaries | | `self-improvement-loops` | Building loops that modify themselves: failure-driven harness self-edits, meta-harness search, evolutionary scaffold search, context mechanism evolution, and acceptance gates for self-modification | | `project-development` | Deciding whether an LLM is appropriate, shaping batch pipelines, creating staged artifacts, or estimating operational cost | | `bdi-mental-states` | Modeling beliefs, desires, intentions, rational action traces, or neuro-symbolic state transformations for agents | ### For Cursor, Codex, and Open Plugins This repository ships as an [Open Plugins](https://open-plugins.com) plugin. Hosts discover skills from the repo-root `skills/` directory (each subdirectory contains a `SKILL.md` file). The manifest lives at `.plugin/plugin.json`. **Cursor (recommended):** 1. Install from the [Cursor Plugin Directory](https://cursor.directory/plugins/context-engineering), or clone this repo and point Cursor at the plugin root. 2. Cursor reads `.plugin/plugin.json` and discovers the repo-root `skills/` directory through the Open Plugins manifest. 3. For project-local manual installs, copy skill directories into `.cursor/skills/`. Do not rely on repository symlinks; they are fragile on Windows and in plugin packaging. **Codex / GitHub Copilot CLI / other Open Plugins hosts:** 1. Clone or add this repository as a plugin directory. 2. The host reads `.plugin/plugin.json` and discovers all published skills under `skills/`. 3. For project-local manual installs, copy skill directories into `.codex/skills/` or the host's documented Agent Skills directory. ### Using Individual Skills Agent Skills require a **directory layout**, not a flat markdown file. Copy the skill folder into your project's skills directory: ```bash # Example: add just the context-fundamentals skill to a Cursor project mkdir -p .cursor/skills cp -R skills/context-fundamentals .cursor/skills/ # Claude Code project-scoped install (same directory layout) mkdir -p .claude/skills cp -R skills/context-fundamentals .claude/skills/ # Codex project-scoped install mkdir -p .codex/skills cp -R skills/context-fundamentals .codex/skills/ # Generic Agent Skills repo-scoped install (Codex/OpenAI, Copilot CLI, Open Plugins hosts) mkdir -p .agents/skills cp -R skills/context-fundamentals .agents/skills/ ``` Do not flatten `SKILL.md` into a single file at `.claude/skills/context-fundamentals.md`. That breaks relative `references/` paths and violates the Agent Skills directory spec used by Cursor, Claude Code, and Codex. Available skills: `context-fundamentals`, `context-degradation`, `context-compression`, `context-optimization`, `latent-briefing`, `multi-agent-patterns`, `long-horizon-prompting`, `memory-systems`, `tool-design`, `filesystem-context`, `hosted-agents`, `evaluation`, `advanced-evaluation`, `harness-engineering`, `self-improvement-loops`, `project-development`, `bdi-mental-states` ### For Custom Implementations Extract the principles and patterns from any skill and implement them in your agent framework. The skills are deliberately platform-agnostic. ## Examples The [examples](examples/) folder contains complete system designs that demonstrate how multiple skills work together in practice. | Example | Description | Skills Applied | |---------|-------------|----------------| | [digital-brain-skill](examples/digital-brain-skill/) | **NEW** Personal operating system for founders and creators. Complete Claude Code skill with 6 modules, 4 automation scripts | context-fundamentals, context-optimization, memory-systems, tool-design, multi-agent-patterns, evaluation, project-development | | [x-to-book-system](examples/x-to-book-system/) | Multi-agent system that monitors X accounts and generates daily synthesized books | multi-agent-patterns, memory-systems, context-optimization, tool-design, evaluation | | [llm-as-judge-skills](examples/llm-as-judge-skills/) | Production-ready LLM evaluation tools with TypeScript implementation, 19 passing tests | advanced-evaluation, tool-design, context-fundamentals, evaluation | | [book-sft-pipeline](examples/book-sft-pipeline/) | Train models to write in any author's style. Includes Gertrude Stein case study with 70% human score on Pangram, $2 total cost | project-development, context-compression, multi-agent-patterns, evaluation | | [interleaved-thinking](examples/interleaved-thinking/) | Reasoning trace optimizer that captures, analyzes, and converts agent failure patterns into generated skills | evaluation, advanced-evaluation, context-degradation, harness-engineering | | [long-horizon-prompt-lab](examples/long-horizon-prompt-lab/) | Production-ready educational website: method guide, copyable task-brief template, four complete prompt rewrites, structural audits, and a caveated research/vendor reference catalog | long-horizon-prompting, harness-engineering, multi-agent-patterns, advanced-evaluation | Each example includes: - Complete PRD with architecture decisions - Skills mapping showing which concepts informed each decision - Implementation guidance ### Digital Brain Skill Example The [digital-brain-skill](examples/digital-brain-skill/) example is a complete personal operating system demonstrating comprehensive skills application: - **Progressive Disclosure**: 3-level loading (SKILL.md → MODULE.md → data files) - **Module Isolation**: 6 independent modules (identity, content, knowledge, network, operations, agents) - **Append-Only Memory**: JSONL files with schema-first lines for agent-friendly parsing - **Automation Scripts**: 4 consolidated tools (weekly_review, content_ideas, stale_contacts, idea_to_draft) Includes detailed traceability in [HOW-SKILLS-BUILT-THIS.md](examples/digital-brain-skill/HOW-SKILLS-BUILT-THIS.md) mapping every architectural decision to specific skill principles. ### LLM-as-Judge Skills Example The [llm-as-judge-skills](examples/llm-as-judge-skills/) example is a complete TypeScript implementation demonstrating: - **Direct Scoring**: Evaluate responses against weighted criteria with rubric support - **Pairwise Comparison**: Compare responses with position bias mitigation - **Rubric Generation**: Create domain-specific evaluation standards - **EvaluatorAgent**: High-level agent combining all evaluation capabilities ### Book SFT Pipeline Example The [book-sft-pipeline](examples/book-sft-pipeline/) example demonstrates training small models (8B) to write in any author's style: - **Intelligent Segmentation**: Two-tier chunking with overlap for maximum training examples - **Prompt Diversity**: 15+ templates to prevent memorization and force style learning - **Tinker Integration**: Complete LoRA training workflow with $2 total cost - **Validation Methodology**: Modern scenario testing proves style transfer vs content memorization Integrates with context engineering skills: project-development, context-compression, multi-agent-patterns, evaluation. ## Researcher Operating System The [researcher](researcher/) directory is a file-based operating system for turning external research into skill changes. It exists so this repository can act as a compounding source of truth instead of an anthology. ### Measured router-benchmark results The skill router (which decides whether the right skill gets loaded for a given task) has been benchmarked end-to-end against four frontier models via the [Cursor SDK](https://cursor.com/docs/sdk/typescript). Three full sweeps (50 prompts x 4 models x 3 replications = 600 calls each): - Baseline: [`researcher/benchmarks/router/results-published/2026-05-15.md`](researcher/benchmarks/router/results-published/2026-05-15.md) - After targeted description rewrites: [`researcher/benchmarks/router/results-published/2026-05-15-v2.md`](researcher/benchmarks/router/results-published/2026-05-15-v2.md) (includes delta-vs-baseline) - After corpus-wide hardening: [`researcher/benchmarks/router/results-published/2026-05-19.md`](researcher/benchmarks/router/results-published/2026-05-19.md) (600/600 usable records, 0 format failures) Per-skill effect size for the three skills the data flagged: | Skill | Baseline top-1 | After rewrite | Delta | | --- | --- | --- | --- | | `context-fundamentals` | 0.255 | 0.489 | +23.4pp | | `project-development` | 0.750 | 1.000 | +25pp (now perfect) | | `tool-design` | 0.729 | 0.807 | +7.8pp | Per-model top-1 accuracy after the corpus-wide hardening pass: | Model | Top-1 | Top-3 | | --- | --- | --- | | gemini-3.1-pro | 0.920 | 0.933 | | composer-2 | 0.913 | 0.947 | | gpt-5.5 | 0.913 | 0.973 | | claude-opus-4-7 | 0.840 | 0.933 | Reproduce any of these numbers exactly via the runner under `researcher/benchmarks/sdk-runner/`. ### What it includes Current counts and compatibility status are generated in the [live corpus inventory](researcher/generated/corpus-summary.md). Dated benchmark and release reports retain their original snapshot counts. - **Source registry** (`researcher/source-registry.md`): priority sources, exclusion rules, monitoring queries. - **Rubrics** (`researcher/rubrics/`): content curation, skill change, harness change, pairwise skill revision. - **Mechanism registry** (`researcher/mechanisms/registry.jsonl` + `ledgers/`): accepted behavior changes used as the primary novelty signal, with append-only accepted/rejected ledgers for institutional memory. - **Claim provenance** (`researcher/claims/index.jsonl`): provenance-tracked claims with source URL, evidence strength, volatility, and last reviewed date. - **Corpus index** (`researcher/corpus/index.json`): canonical machine-readable map of skills, activation scenarios, mechanism IDs, and claim IDs. - **Run state machine** (`researcher/runs//run-state.json`): `initialized -> retrieved -> evaluated -> proposed -> novelty_checked -> validated -> pr_ready -> closed`. - **Activation regression tests** (`researcher/fixtures/activation-cases.jsonl`): deterministic prompts that catch skill-boundary confusion. - **Adversarial benchmark harness** (`researcher/benchmarks/`): scenarios that try to game the loop (duplicate mechanisms, unretrieved evidence, wrong rubric math, self-approved rubric changes, weak-evidence novelty). - **Continuous loop** (`researcher/scripts/loop_*.py` + `researcher/orchestration/launchd/`): inbox, source discovery, one-state-at-a-time advancement, daily ops, parked review queue, launchd service definitions. - **Skill health gate** (`researcher/scripts/skill_health.py`): deterministic body-quality scoring. Published scores remain dated evidence; local runs produce ignored runtime reports. - **Public export boundary** (`governance/export-policy.yaml` + `researcher/scripts/validate_export.py`): allowlisted projections from private or restricted records into reviewable public staging trees without publishing private source locators or digests. - **Schema and artifact contract** (`researcher/schemas/` + `researcher/scripts/artifact_store.py`): digest-pinned cross-runtime schemas, typed IDs, private bindings, exact-byte CAS, candidate freeze receipts, and Python/TypeScript conformance. ### Operator commands Install the validation dependencies once before running local gates: ```bash python3 -m pip install -r requirements-dev.txt ``` ```bash # Deterministic gates (also run in CI on every PR) python3 -m unittest researcher.scripts.tests.test_skill_frontmatter python3 researcher/scripts/validate_platform_compat.py --require-reference-validator python3 researcher/scripts/validate_repo.py --strict python3 researcher/scripts/skill_health.py --strict --no-history python3 researcher/scripts/run_benchmarks.py python3 researcher/scripts/check_activation_cases.py # Per-run readiness (active runs only) python3 researcher/scripts/validate_run.py --run-dir researcher/runs/ # Continuous loop, manual python3 researcher/scripts/loop_discover.py python3 researcher/scripts/loop_step.py --allow-fetch python3 researcher/scripts/loop_daily.py python3 researcher/scripts/loop_status.py # Continuous loop, daemon (macOS) researcher/orchestration/launchd/install.sh # install launchd jobs (10-min step, 12h discover, daily ops) researcher/orchestration/launchd/uninstall.sh # remove launchd jobs ``` See [researcher/runbooks/continuous-operation.md](researcher/runbooks/continuous-operation.md) for daemon details, budgets, and the human review surface. ### Guarantees - The loop never invokes paid LLMs or makes outbound writes; HTTP retrieval is stdlib-only with a 1.5 MB cap and a 30-second timeout. - Mechanism promotion requires a recorded human reviewer and a passing run-readiness check. - All queue mutations are atomic (temp file + `os.replace`) and serialized via `fcntl` locks. - Agents may prepare PRs after gates pass; merge and push remain human-controlled. ## Star History ## Structure Each skill follows the Agent Skills specification: ``` skill-name/ ├── SKILL.md # Required: instructions + metadata ├── scripts/ # Optional: executable code demonstrating concepts └── references/ # Optional: additional documentation and resources ``` See the [template](template/) folder for the canonical skill structure. ## Contributing This repository follows the Agent Skills open development model. Contributions are welcome from the broader ecosystem. When contributing: 1. Follow the skill template structure 2. Provide clear, actionable instructions 3. Include working examples where appropriate 4. Document trade-offs and potential issues 5. Keep SKILL.md under 500 lines for optimal performance Feel free to contact [Muratcan Koylan](https://x.com/muratcan) for collaboration opportunities or any inquiries. ## License MIT License - see LICENSE file for details. ## References The principles in these skills are derived from research and production experience at leading AI labs and framework developers. Each skill includes references to the underlying research and case studies that inform its recommendations. --- ## File: docs/agentskills.md --- name: agent-skills-format doc_type: reference source_url: No --- Overview Copy page A simple, open format for giving agents new capabilities and expertise. Agent Skills are folders of instructions, scripts, and resources that agents can discover and use to do things more accurately and efficiently. ​ Why Agent Skills? Agents are increasingly capable, but often don’t have the context they need to do real work reliably. Skills solve this by giving agents access to procedural knowledge and company-, team-, and user-specific context they can load on demand. Agents with access to a set of skills can extend their capabilities based on the task they’re working on. For skill authors: Build capabilities once and deploy them across multiple agent products. For compatible agents: Support for skills lets end users give agents new capabilities out of the box. For teams and enterprises: Capture organizational knowledge in portable, version-controlled packages. ​ What can Agent Skills enable? Domain expertise: Package specialized knowledge into reusable instructions, from legal review processes to data analysis pipelines. New capabilities: Give agents new capabilities (e.g. creating presentations, building MCP servers, analyzing datasets). Repeatable workflows: Turn multi-step tasks into consistent and auditable workflows. Interoperability: Reuse the same skill across different skills-compatible agent products. ​ Adoption Agent Skills are supported by leading AI development tools. OpenCode Cursor Amp Letta Goose GitHub VS Code Claude Code Claude OpenAI Codex ​ Open development The Agent Skills format was originally developed by Anthropic, released as an open standard, and has been adopted by a growing number of agent products. The standard is open to contributions from the broader ecosystem. What are skills? Copy page Agent Skills are a lightweight, open format for extending AI agent capabilities with specialized knowledge and workflows. At its core, a skill is a folder containing a SKILL.md file. This file includes metadata (name and description, at minimum) and instructions that tell an agent how to perform a specific task. Skills can also bundle scripts, templates, and reference materials. my-skill/ ├── SKILL.md # Required: instructions + metadata ├── scripts/ # Optional: executable code ├── references/ # Optional: documentation └── assets/ # Optional: templates, resources ​ How skills work Skills use progressive disclosure to manage context efficiently: Discovery: At startup, agents load only the name and description of each available skill, just enough to know when it might be relevant. Activation: When a task matches a skill’s description, the agent reads the full SKILL.md instructions into context. Execution: The agent follows the instructions, optionally loading referenced files or executing bundled code as needed. This approach keeps agents fast while giving them access to more context on demand. ​ The SKILL.md file Every skill starts with a SKILL.md file containing YAML frontmatter and Markdown instructions: --- name: pdf-processing --- # PDF Processing ## When to use this skill Use this skill when the user needs to work with PDF files... ## How to extract text 1. Use pdfplumber for text extraction... ## How to fill forms ... The following frontmatter is required at the top of SKILL.md: name: A short identifier The Markdown body contains the actual instructions and has no specific restrictions on structure or content. This simple format has some key advantages: Self-documenting: A skill author or user can read a SKILL.md and understand what it does, making skills easy to audit and improve. Extensible: Skills can range in complexity from just text instructions to executable code, assets, and templates. Portable: Skills are just files, so they’re easy to edit, version, and share. ​ Next steps View the specification to understand the full format. Add skills support to your agent to build a compatible client. See example skills on GitHub. Read authoring best practices for writing effective skills. Use the reference library to validate skills and generate prompt XML. Specification Copy page The complete format specification for Agent Skills. This document defines the Agent Skills format. ​ Directory structure A skill is a directory containing at minimum a SKILL.md file: skill-name/ └── SKILL.md # Required You can optionally include additional directories such as scripts/, references/, and assets/ to support your skill. ​ SKILL.md format The SKILL.md file must contain YAML frontmatter followed by Markdown content. ​ Frontmatter (required) --- name: skill-name --- With optional fields: --- name: pdf-processing license: Apache-2.0 metadata: author: example-org version: "1.0" --- Field Required Constraints name Yes Max 64 characters. Lowercase letters, numbers, and hyphens only. Must not start or end with a hyphen. description Yes Max 1024 characters. Non-empty. Describes what the skill does and when to use it. license No License name or reference to a bundled license file. compatibility No Max 500 characters. Indicates environment requirements (intended product, system packages, network access, etc.). metadata No Arbitrary key-value mapping for additional metadata. allowed-tools No Space-delimited list of pre-approved tools the skill may use. (Experimental) ​ name field The required name field: Must be 1-64 characters May only contain unicode lowercase alphanumeric characters and hyphens (a-z and -) Must not start or end with - Must not contain consecutive hyphens (--) Must match the parent directory name Valid examples: name: pdf-processing name: data-analysis name: code-review Invalid examples: name: PDF-Processing # uppercase not allowed name: -pdf # cannot start with hyphen name: pdf--processing # consecutive hyphens not allowed ​ description field The required description field: Must be 1-1024 characters Should describe both what the skill does and when to use it Should include specific keywords that help agents identify relevant tasks Good example: Poor example: ​ license field The optional license field: Specifies the license applied to the skill We recommend keeping it short (either the name of a license or the name of a bundled license file) Example: license: Proprietary. LICENSE.txt has complete terms ​ compatibility field The optional compatibility field: Must be 1-500 characters if provided Should only be included if your skill has specific environment requirements Can indicate intended product, required system packages, network access needs, etc. Examples: compatibility: Designed for Claude Code (or similar products) compatibility: Requires git, docker, jq, and access to the internet Most skills do not need the compatibility field. ​ metadata field The optional metadata field: A map from string keys to string values Clients can use this to store additional properties not defined by the Agent Skills spec We recommend making your key names reasonably unique to avoid accidental conflicts Example: metadata: author: example-org version: "1.0" ​ allowed-tools field The optional allowed-tools field: A space-delimited list of tools that are pre-approved to run Experimental. Support for this field may vary between agent implementations Example: allowed-tools: Bash(git:*) Bash(jq:*) Read ​ Body content The Markdown body after the frontmatter contains the skill instructions. There are no format restrictions. Write whatever helps agents perform the task effectively. Recommended sections: Step-by-step instructions Examples of inputs and outputs Common edge cases Note that the agent will load this entire file once it’s decided to activate a skill. Consider splitting longer SKILL.md content into referenced files. ​ Optional directories ​ scripts/ Contains executable code that agents can run. Scripts should: Be self-contained or clearly document dependencies Include helpful error messages Handle edge cases gracefully Supported languages depend on the agent implementation. Common options include Python, Bash, and JavaScript. ​ references/ Contains additional documentation that agents can read when needed: REFERENCE.md - Detailed technical reference FORMS.md - Form templates or structured data formats Domain-specific files (finance.md, legal.md, etc.) Keep individual reference files focused. Agents load these on demand, so smaller files mean less use of context. ​ assets/ Contains static resources: Templates (document templates, configuration templates) Images (diagrams, examples) Data files (lookup tables, schemas) ​ Progressive disclosure Skills should be structured for efficient use of context: Metadata (~100 tokens): The name and description fields are loaded at startup for all skills Instructions (< 5000 tokens recommended): The full SKILL.md body is loaded when the skill is activated Resources (as needed): Files (e.g. those in scripts/, references/, or assets/) are loaded only when required Keep your main SKILL.md under 500 lines. Move detailed reference material to separate files. ​ File references When referencing other files in your skill, use relative paths from the skill root: See [the reference guide](references/REFERENCE.md) for details. Run the extraction script: scripts/extract.py Keep file references one level deep from SKILL.md. Avoid deeply nested reference chains. ​ Validation Use the skills-ref reference library to validate your skills: skills-ref validate ./my-skill This checks that your SKILL.md frontmatter is valid and follows all naming conventions. Integrate skills into your agent Copy page How to add Agent Skills support to your agent or tool. This guide explains how to add skills support to an AI agent or development tool. ​ Integration approaches The two main approaches to integrating skills are: Filesystem-based agents operate within a computer environment (bash/unix) and represent the most capable option. Skills are activated when models issue shell commands like cat /path/to/my-skill/SKILL.md. Bundled resources are accessed through shell commands. Tool-based agents function without a dedicated computer environment. Instead, they implement tools allowing models to trigger skills and access bundled assets. The specific tool implementation is up to the developer. ​ Overview A skills-compatible agent needs to: Discover skills in configured directories Load metadata (name and description) at startup Match user tasks to relevant skills Activate skills by loading full instructions Execute scripts and access resources as needed ​ Skill discovery Skills are folders containing a SKILL.md file. Your agent should scan configured directories for valid skills. ​ Loading metadata At startup, parse only the frontmatter of each SKILL.md file. This keeps initial context usage low. ​ Parsing frontmatter function parseMetadata(skillPath): content = readFile(skillPath + "/SKILL.md") frontmatter = extractYAMLFrontmatter(content) return { name: frontmatter.name, path: skillPath } ​ Injecting into context Include skill metadata in the system prompt so the model knows what skills are available. Follow your platform’s guidance for system prompt updates. For example, for Claude models, the recommended format uses XML: pdf-processing Extracts text and tables from PDF files, fills forms, merges documents. /path/to/skills/pdf-processing/SKILL.md data-analysis Analyzes datasets, generates charts, and creates summary reports. /path/to/skills/data-analysis/SKILL.md For filesystem-based agents, include the location field with the absolute path to the SKILL.md file. For tool-based agents, the location can be omitted. Keep metadata concise. Each skill should add roughly 50-100 tokens to the context. ​ Security considerations Script execution introduces security risks. Consider: Sandboxing: Run scripts in isolated environments Allowlisting: Only execute scripts from trusted skills Confirmation: Ask users before running potentially dangerous operations Logging: Record all script executions for auditing ​ Reference implementation The skills-ref library provides Python utilities and a CLI for working with skills. For example: Validate a skill directory: skills-ref validate Generate XML for agent prompts: skills-ref to-prompt ... Use the library source code as a reference implementation. Skill authoring best practices Copy page Learn how to write effective Skills that Claude can discover and use successfully. Good Skills are concise, well-structured, and tested with real usage. This guide provides practical authoring decisions to help you write Skills that Claude can discover and use effectively. For conceptual background on how Skills work, see the Skills overview. Core principles Concise is key The context window is a public good. Your Skill shares the context window with everything else Claude needs to know, including: The system prompt Conversation history Other Skills' metadata Your actual request Not every token in your Skill has an immediate cost. At startup, only the metadata (name and description) from all Skills is pre-loaded. Claude reads SKILL.md only when the Skill becomes relevant, and reads additional files only as needed. However, being concise in SKILL.md still matters: once Claude loads it, every token competes with conversation history and other context. Default assumption: Claude is already very smart Only add context Claude doesn't already have. Challenge each piece of information: "Does Claude really need this explanation?" "Can I assume Claude knows this?" "Does this paragraph justify its token cost?" Good example: Concise (approximately 50 tokens): ## Extract PDF text Use pdfplumber for text extraction: ```python import pdfplumber with pdfplumber.open("file.pdf") as pdf: text = pdf.pages[0].extract_text() ``` Bad example: Too verbose (approximately 150 tokens): ## Extract PDF text PDF (Portable Document Format) files are a common file format that contains text, images, and other content. To extract text from a PDF, you'll need to use a library. There are many libraries available for PDF processing, but we recommend pdfplumber because it's easy to use and handles most cases well. First, you'll need to install it using pip. Then you can use the code below... The concise version assumes Claude knows what PDFs are and how libraries work. Set appropriate degrees of freedom Match the level of specificity to the task's fragility and variability. High freedom (text-based instructions): Use when: Multiple approaches are valid Decisions depend on context Heuristics guide the approach Example: ## Code review process 1. Analyze the code structure and organization 2. Check for potential bugs or edge cases 3. Suggest improvements for readability and maintainability 4. Verify adherence to project conventions Medium freedom (pseudocode or scripts with parameters): Use when: A preferred pattern exists Some variation is acceptable Configuration affects behavior Example: ## Generate report Use this template and customize as needed: ```python def generate_report(data, format="markdown", include_charts=True): # Process data # Generate output in specified format # Optionally include visualizations ``` Low freedom (specific scripts, few or no parameters): Use when: Operations are fragile and error-prone Consistency is critical A specific sequence must be followed Example: ## Database migration Run exactly this script: ```bash python scripts/migrate.py --verify --backup ``` Do not modify the command or add additional flags. Analogy: Think of Claude as a robot exploring a path: Narrow bridge with cliffs on both sides: There's only one safe way forward. Provide specific guardrails and exact instructions (low freedom). Example: database migrations that must run in exact sequence. Open field with no hazards: Many paths lead to success. Give general direction and trust Claude to find the best route (high freedom). Example: code reviews where context determines the best approach. Test with all models you plan to use Skills act as additions to models, so effectiveness depends on the underlying model. Test your Skill with all the models you plan to use it with. Testing considerations by model: Claude Haiku (fast, economical): Does the Skill provide enough guidance? Claude Sonnet (balanced): Is the Skill clear and efficient? Claude Opus (powerful reasoning): Does the Skill avoid over-explaining? What works perfectly for Opus might need more detail for Haiku. If you plan to use your Skill across multiple models, aim for instructions that work well with all of them. Skill structure YAML Frontmatter: The SKILL.md frontmatter requires two fields: name: Maximum 64 characters Must contain only lowercase letters, numbers, and hyphens Cannot contain XML tags Cannot contain reserved words: "anthropic", "claude" Must be non-empty Maximum 1024 characters Cannot contain XML tags Should describe what the Skill does and when to use it For complete Skill structure details, see the Skills overview. Naming conventions Use consistent naming patterns to make Skills easier to reference and discuss. We recommend using gerund form (verb + -ing) for Skill names, as this clearly describes the activity or capability the Skill provides. Remember that the name field must use lowercase letters, numbers, and hyphens only. Good naming examples (gerund form): processing-pdfs analyzing-spreadsheets managing-databases testing-code writing-documentation Acceptable alternatives: Noun phrases: pdf-processing, spreadsheet-analysis Action-oriented: process-pdfs, analyze-spreadsheets Avoid: Vague names: helper, utils, tools Overly generic: documents, data, files Reserved words: anthropic-helper, claude-tools Inconsistent patterns within your skill collection Consistent naming makes it easier to: Reference Skills in documentation and conversations Understand what a Skill does at a glance Organize and search through multiple Skills Maintain a professional, cohesive skill library Writing effective descriptions The description field enables Skill discovery and should include both what the Skill does and when to use it. Always write in third person. The description is injected into the system prompt, and inconsistent point-of-view can cause discovery problems. Good: "Processes Excel files and generates reports" Avoid: "I can help you process Excel files" Avoid: "You can use this to process Excel files" Be specific and include key terms. Include both what the Skill does and specific triggers/contexts for when to use it. Each Skill has exactly one description field. The description is critical for skill selection: Claude uses it to choose the right Skill from potentially 100+ available Skills. Your description must provide enough detail for Claude to know when to select this Skill, while the rest of SKILL.md provides the implementation details. Effective examples: PDF Processing skill: Excel Analysis skill: Git Commit Helper skill: Avoid vague descriptions like these: Progressive disclosure patterns SKILL.md serves as an overview that points Claude to detailed materials as needed, like a table of contents in an onboarding guide. For an explanation of how progressive disclosure works, see How Skills work in the overview. Practical guidance: Keep SKILL.md body under 500 lines for optimal performance Split content into separate files when approaching this limit Use the patterns below to organize instructions, code, and resources effectively Visual overview: From simple to complex A basic Skill starts with just a SKILL.md file containing metadata and instructions: Simple SKILL.md file showing YAML frontmatter and markdown body As your Skill grows, you can bundle additional content that Claude loads only when needed: Bundling additional reference files like reference.md and forms.md. The complete Skill directory structure might look like this: pdf/ ├── SKILL.md # Main instructions (loaded when triggered) ├── FORMS.md # Form-filling guide (loaded as needed) ├── reference.md # API reference (loaded as needed) ├── examples.md # Usage examples (loaded as needed) └── scripts/ ├── analyze_form.py # Utility script (executed, not loaded) ├── fill_form.py # Form filling script └── validate.py # Validation script Pattern 1: High-level guide with references --- name: pdf-processing --- # PDF Processing ## Quick start Extract text with pdfplumber: ```python import pdfplumber with pdfplumber.open("file.pdf") as pdf: text = pdf.pages[0].extract_text() ``` ## Advanced features **Form filling**: See [FORMS.md](FORMS.md) for complete guide **API reference**: See [REFERENCE.md](REFERENCE.md) for all methods **Examples**: See [EXAMPLES.md](EXAMPLES.md) for common patterns Claude loads FORMS.md, REFERENCE.md, or EXAMPLES.md only when needed. Pattern 2: Domain-specific organization For Skills with multiple domains, organize content by domain to avoid loading irrelevant context. When a user asks about sales metrics, Claude only needs to read sales-related schemas, not finance or marketing data. This keeps token usage low and context focused. bigquery-skill/ ├── SKILL.md (overview and navigation) └── reference/ ├── finance.md (revenue, billing metrics) ├── sales.md (opportunities, pipeline) ├── product.md (API usage, features) └── marketing.md (campaigns, attribution) SKILL.md # BigQuery Data Analysis ## Available datasets **Finance**: Revenue, ARR, billing → See [reference/finance.md](reference/finance.md) **Sales**: Opportunities, pipeline, accounts → See [reference/sales.md](reference/sales.md) **Product**: API usage, features, adoption → See [reference/product.md](reference/product.md) **Marketing**: Campaigns, attribution, email → See [reference/marketing.md](reference/marketing.md) ## Quick search Find specific metrics using grep: ```bash grep -i "revenue" reference/finance.md grep -i "pipeline" reference/sales.md grep -i "api usage" reference/product.md ``` Pattern 3: Conditional details Show basic content, link to advanced content: # DOCX Processing ## Creating documents Use docx-js for new documents. See [DOCX-JS.md](DOCX-JS.md). ## Editing documents For simple edits, modify the XML directly. **For tracked changes**: See [REDLINING.md](REDLINING.md) **For OOXML details**: See [OOXML.md](OOXML.md) Claude reads REDLINING.md or OOXML.md only when the user needs those features. Avoid deeply nested references Claude may partially read files when they're referenced from other referenced files. When encountering nested references, Claude might use commands like head -100 to preview content rather than reading entire files, resulting in incomplete information. Keep references one level deep from SKILL.md. All reference files should link directly from SKILL.md to ensure Claude reads complete files when needed. Bad example: Too deep: # SKILL.md See [advanced.md](advanced.md)... # advanced.md See [details.md](details.md)... # details.md Here's the actual information... Good example: One level deep: # SKILL.md **Basic usage**: [instructions in SKILL.md] **Advanced features**: See [advanced.md](advanced.md) **API reference**: See [reference.md](reference.md) **Examples**: See [examples.md](examples.md) Structure longer reference files with table of contents For reference files longer than 100 lines, include a table of contents at the top. This ensures Claude can see the full scope of available information even when previewing with partial reads. Example: # API Reference ## Contents - Authentication and setup - Core methods (create, read, update, delete) - Advanced features (batch operations, webhooks) - Error handling patterns - Code examples ## Authentication and setup ... ## Core methods ... Claude can then read the complete file or jump to specific sections as needed. For details on how this filesystem-based architecture enables progressive disclosure, see the Runtime environment section in the Advanced section below. Workflows and feedback loops Use workflows for complex tasks Break complex operations into clear, sequential steps. For particularly complex workflows, provide a checklist that Claude can copy into its response and check off as it progresses. Example 1: Research synthesis workflow (for Skills without code): ## Research synthesis workflow Copy this checklist and track your progress: ``` Research Progress: - [ ] Step 1: Read all source documents - [ ] Step 2: Identify key themes - [ ] Step 3: Cross-reference claims - [ ] Step 4: Create structured summary - [ ] Step 5: Verify citations ``` **Step 1: Read all source documents** Review each document in the `sources/` directory. Note the main arguments and supporting evidence. **Step 2: Identify key themes** Look for patterns across sources. What themes appear repeatedly? Where do sources agree or disagree? **Step 3: Cross-reference claims** For each major claim, verify it appears in the source material. Note which source supports each point. **Step 4: Create structured summary** Organize findings by theme. Include: - Main claim - Supporting evidence from sources - Conflicting viewpoints (if any) **Step 5: Verify citations** Check that every claim references the correct source document. If citations are incomplete, return to Step 3. This example shows how workflows apply to analysis tasks that don't require code. The checklist pattern works for any complex, multi-step process. Example 2: PDF form filling workflow (for Skills with code): ## PDF form filling workflow Copy this checklist and check off items as you complete them: ``` Task Progress: - [ ] Step 1: Analyze the form (run analyze_form.py) - [ ] Step 2: Create field mapping (edit fields.json) - [ ] Step 3: Validate mapping (run validate_fields.py) - [ ] Step 4: Fill the form (run fill_form.py) - [ ] Step 5: Verify output (run verify_output.py) ``` **Step 1: Analyze the form** Run: `python scripts/analyze_form.py input.pdf` This extracts form fields and their locations, saving to `fields.json`. **Step 2: Create field mapping** Edit `fields.json` to add values for each field. **Step 3: Validate mapping** Run: `python scripts/validate_fields.py fields.json` Fix any validation errors before continuing. **Step 4: Fill the form** Run: `python scripts/fill_form.py input.pdf fields.json output.pdf` **Step 5: Verify output** Run: `python scripts/verify_output.py output.pdf` If verification fails, return to Step 2. Clear steps prevent Claude from skipping critical validation. The checklist helps both Claude and you track progress through multi-step workflows. Implement feedback loops Common pattern: Run validator → fix errors → repeat This pattern greatly improves output quality. Example 1: Style guide compliance (for Skills without code): ## Content review process 1. Draft your content following the guidelines in STYLE_GUIDE.md 2. Review against the checklist: - Check terminology consistency - Verify examples follow the standard format - Confirm all required sections are present 3. If issues found: - Note each issue with specific section reference - Revise the content - Review the checklist again 4. Only proceed when all requirements are met 5. Finalize and save the document This shows the validation loop pattern using reference documents instead of scripts. The "validator" is STYLE_GUIDE.md, and Claude performs the check by reading and comparing. Example 2: Document editing process (for Skills with code): ## Document editing process 1. Make your edits to `word/document.xml` 2. **Validate immediately**: `python ooxml/scripts/validate.py unpacked_dir/` 3. If validation fails: - Review the error message carefully - Fix the issues in the XML - Run validation again 4. **Only proceed when validation passes** 5. Rebuild: `python ooxml/scripts/pack.py unpacked_dir/ output.docx` 6. Test the output document The validation loop catches errors early. Content guidelines Avoid time-sensitive information Don't include information that will become outdated: Bad example: Time-sensitive (will become wrong): If you're doing this before August 2025, use the old API. After August 2025, use the new API. Good example (use "old patterns" section): ## Current method Use the v2 API endpoint: `api.example.com/v2/messages` ## Old patterns Legacy v1 API (deprecated 2025-08) The v1 API used: `api.example.com/v1/messages` This endpoint is no longer supported. The old patterns section provides historical context without cluttering the main content. Use consistent terminology Choose one term and use it throughout the Skill: Good - Consistent: Always "API endpoint" Always "field" Always "extract" Bad - Inconsistent: Mix "API endpoint", "URL", "API route", "path" Mix "field", "box", "element", "control" Mix "extract", "pull", "get", "retrieve" Consistency helps Claude understand and follow instructions. Common patterns Template pattern Provide templates for output format. Match the level of strictness to your needs. For strict requirements (like API responses or data formats): ## Report structure ALWAYS use this exact template structure: ```markdown # [Analysis Title] ## Executive summary [One-paragraph overview of key findings] ## Key findings - Finding 1 with supporting data - Finding 2 with supporting data - Finding 3 with supporting data ## Recommendations 1. Specific actionable recommendation 2. Specific actionable recommendation ``` For flexible guidance (when adaptation is useful): ## Report structure Here is a sensible default format, but use your best judgment based on the analysis: ```markdown # [Analysis Title] ## Executive summary [Overview] ## Key findings [Adapt sections based on what you discover] ## Recommendations [Tailor to the specific context] ``` Adjust sections as needed for the specific analysis type. Examples pattern For Skills where output quality depends on seeing examples, provide input/output pairs just like in regular prompting: ## Commit message format Generate commit messages following these examples: **Example 1:** Input: Added user authentication with JWT tokens Output: ``` feat(auth): implement JWT-based authentication Add login endpoint and token validation middleware ``` **Example 2:** Input: Fixed bug where dates displayed incorrectly in reports Output: ``` fix(reports): correct date formatting in timezone conversion Use UTC timestamps consistently across report generation ``` **Example 3:** Input: Updated dependencies and refactored error handling Output: ``` chore: update dependencies and refactor error handling - Upgrade lodash to 4.17.21 - Standardize error response format across endpoints ``` Follow this style: type(scope): brief description, then detailed explanation. Examples help Claude understand the desired style and level of detail more clearly than descriptions alone. Conditional workflow pattern Guide Claude through decision points: ## Document modification workflow 1. Determine the modification type: **Creating new content?** → Follow "Creation workflow" below **Editing existing content?** → Follow "Editing workflow" below 2. Creation workflow: - Use docx-js library - Build document from scratch - Export to .docx format 3. Editing workflow: - Unpack existing document - Modify XML directly - Validate after each change - Repack when complete If workflows become large or complicated with many steps, consider pushing them into separate files and tell Claude to read the appropriate file based on the task at hand. Evaluation and iteration Build evaluations first Create evaluations BEFORE writing extensive documentation. This ensures your Skill solves real problems rather than documenting imagined ones. Evaluation-driven development: Identify gaps: Run Claude on representative tasks without a Skill. Document specific failures or missing context Create evaluations: Build three scenarios that test these gaps Establish baseline: Measure Claude's performance without the Skill Write minimal instructions: Create just enough content to address the gaps and pass evaluations Iterate: Execute evaluations, compare against baseline, and refine This approach ensures you're solving actual problems rather than anticipating requirements that may never materialize. Evaluation structure: { "skills": ["pdf-processing"], "query": "Extract all text from this PDF file and save it to output.txt", "files": ["test-files/document.pdf"], "expected_behavior": [ "Successfully reads the PDF file using an appropriate PDF processing library or command-line tool", "Extracts text content from all pages in the document without missing any pages", "Saves the extracted text to a file named output.txt in a clear, readable format" ] } This example demonstrates a data-driven evaluation with a simple testing rubric. We do not currently provide a built-in way to run these evaluations. Users can create their own evaluation system. Evaluations are your source of truth for measuring Skill effectiveness. Develop Skills iteratively with Claude The most effective Skill development process involves Claude itself. Work with one instance of Claude ("Claude A") to create a Skill that will be used by other instances ("Claude B"). Claude A helps you design and refine instructions, while Claude B tests them in real tasks. This works because Claude models understand both how to write effective agent instructions and what information agents need. Creating a new Skill: Complete a task without a Skill: Work through a problem with Claude A using normal prompting. As you work, you'll naturally provide context, explain preferences, and share procedural knowledge. Notice what information you repeatedly provide. Identify the reusable pattern: After completing the task, identify what context you provided that would be useful for similar future tasks. Example: If you worked through a BigQuery analysis, you might have provided table names, field definitions, filtering rules (like "always exclude test accounts"), and common query patterns. Ask Claude A to create a Skill: "Create a Skill that captures this BigQuery analysis pattern we just used. Include the table schemas, naming conventions, and the rule about filtering test accounts." Claude models understand the Skill format and structure natively. You don't need special system prompts or a "writing skills" skill to get Claude to help create Skills. Simply ask Claude to create a Skill and it will generate properly structured SKILL.md content with appropriate frontmatter and body content. Review for conciseness: Check that Claude A hasn't added unnecessary explanations. Ask: "Remove the explanation about what win rate means - Claude already knows that." Improve information architecture: Ask Claude A to organize the content more effectively. For example: "Organize this so the table schema is in a separate reference file. We might add more tables later." Test on similar tasks: Use the Skill with Claude B (a fresh instance with the Skill loaded) on related use cases. Observe whether Claude B finds the right information, applies rules correctly, and handles the task successfully. Iterate based on observation: If Claude B struggles or misses something, return to Claude A with specifics: "When Claude used this Skill, it forgot to filter by date for Q4. Should we add a section about date filtering patterns?" Iterating on existing Skills: The same hierarchical pattern continues when improving Skills. You alternate between: Working with Claude A (the expert who helps refine the Skill) Testing with Claude B (the agent using the Skill to perform real work) Observing Claude B's behavior and bringing insights back to Claude A Use the Skill in real workflows: Give Claude B (with the Skill loaded) actual tasks, not test scenarios Observe Claude B's behavior: Note where it struggles, succeeds, or makes unexpected choices Example observation: "When I asked Claude B for a regional sales report, it wrote the query but forgot to filter out test accounts, even though the Skill mentions this rule." Return to Claude A for improvements: Share the current SKILL.md and describe what you observed. Ask: "I noticed Claude B forgot to filter test accounts when I asked for a regional report. The Skill mentions filtering, but maybe it's not prominent enough?" Review Claude A's suggestions: Claude A might suggest reorganizing to make rules more prominent, using stronger language like "MUST filter" instead of "always filter", or restructuring the workflow section. Apply and test changes: Update the Skill with Claude A's refinements, then test again with Claude B on similar requests Repeat based on usage: Continue this observe-refine-test cycle as you encounter new scenarios. Each iteration improves the Skill based on real agent behavior, not assumptions. Gathering team feedback: Share Skills with teammates and observe their usage Ask: Does the Skill activate when expected? Are instructions clear? What's missing? Incorporate feedback to address blind spots in your own usage patterns Why this approach works: Claude A understands agent needs, you provide domain expertise, Claude B reveals gaps through real usage, and iterative refinement improves Skills based on observed behavior rather than assumptions. Observe how Claude navigates Skills As you iterate on Skills, pay attention to how Claude actually uses them in practice. Watch for: Unexpected exploration paths: Does Claude read files in an order you didn't anticipate? This might indicate your structure isn't as intuitive as you thought Missed connections: Does Claude fail to follow references to important files? Your links might need to be more explicit or prominent Overreliance on certain sections: If Claude repeatedly reads the same file, consider whether that content should be in the main SKILL.md instead Ignored content: If Claude never accesses a bundled file, it might be unnecessary or poorly signaled in the main instructions Iterate based on these observations rather than assumptions. The 'name' and 'description' in your Skill's metadata are particularly critical. Claude uses these when deciding whether to trigger the Skill in response to the current task. Make sure they clearly describe what the Skill does and when it should be used. Anti-patterns to avoid Avoid Windows-style paths Always use forward slashes in file paths, even on Windows: ✓ Good: scripts/helper.py, reference/guide.md ✗ Avoid: scripts\helper.py, reference\guide.md Unix-style paths work across all platforms, while Windows-style paths cause errors on Unix systems. Avoid offering too many options Don't present multiple approaches unless necessary: **Bad example: Too many choices** (confusing): "You can use pypdf, or pdfplumber, or PyMuPDF, or pdf2image, or..." **Good example: Provide a default** (with escape hatch): "Use pdfplumber for text extraction: ```python import pdfplumber ``` For scanned PDFs requiring OCR, use pdf2image with pytesseract instead." Advanced: Skills with executable code The sections below focus on Skills that include executable scripts. If your Skill uses only markdown instructions, skip to Checklist for effective Skills. Solve, don't punt When writing scripts for Skills, handle error conditions rather than punting to Claude. Good example: Handle errors explicitly: def process_file(path): """Process a file, creating it if it doesn't exist.""" try: with open(path) as f: return f.read() except FileNotFoundError: # Create file with default content instead of failing print(f"File {path} not found, creating default") with open(path, 'w') as f: f.write('') return '' except PermissionError: # Provide alternative instead of failing print(f"Cannot access {path}, using default") return '' Bad example: Punt to Claude: def process_file(path): # Just fail and let Claude figure it out return open(path).read() Configuration parameters should also be justified and documented to avoid "voodoo constants" (Ousterhout's law). If you don't know the right value, how will Claude determine it? Good example: Self-documenting: # HTTP requests typically complete within 30 seconds # Longer timeout accounts for slow connections REQUEST_TIMEOUT = 30 # Three retries balances reliability vs speed # Most intermittent failures resolve by the second retry MAX_RETRIES = 3 Bad example: Magic numbers: TIMEOUT = 47 # Why 47? RETRIES = 5 # Why 5? Provide utility scripts Even if Claude could write a script, pre-made scripts offer advantages: Benefits of utility scripts: More reliable than generated code Save tokens (no need to include code in context) Save time (no code generation required) Ensure consistency across uses Bundling executable scripts alongside instruction files The diagram above shows how executable scripts work alongside instruction files. The instruction file (forms.md) references the script, and Claude can execute it without loading its contents into context. Important distinction: Make clear in your instructions whether Claude should: Execute the script (most common): "Run analyze_form.py to extract fields" Read it as reference (for complex logic): "See analyze_form.py for the field extraction algorithm" For most utility scripts, execution is preferred because it's more reliable and efficient. See the Runtime environment section below for details on how script execution works. Example: ## Utility scripts **analyze_form.py**: Extract all form fields from PDF ```bash python scripts/analyze_form.py input.pdf > fields.json ``` Output format: ```json { "field_name": {"type": "text", "x": 100, "y": 200}, "signature": {"type": "sig", "x": 150, "y": 500} } ``` **validate_boxes.py**: Check for overlapping bounding boxes ```bash python scripts/validate_boxes.py fields.json # Returns: "OK" or lists conflicts ``` **fill_form.py**: Apply field values to PDF ```bash python scripts/fill_form.py input.pdf fields.json output.pdf ``` Use visual analysis When inputs can be rendered as images, have Claude analyze them: ## Form layout analysis 1. Convert PDF to images: ```bash python scripts/pdf_to_images.py form.pdf ``` 2. Analyze each page image to identify form fields 3. Claude can see field locations and types visually In this example, you'd need to write the pdf_to_images.py script. Claude's vision capabilities help understand layouts and structures. Create verifiable intermediate outputs When Claude performs complex, open-ended tasks, it can make mistakes. The "plan-validate-execute" pattern catches errors early by having Claude first create a plan in a structured format, then validate that plan with a script before executing it. Example: Imagine asking Claude to update 50 form fields in a PDF based on a spreadsheet. Without validation, Claude might reference non-existent fields, create conflicting values, miss required fields, or apply updates incorrectly. Solution: Use the workflow pattern shown above (PDF form filling), but add an intermediate changes.json file that gets validated before applying changes. The workflow becomes: analyze → create plan file → validate plan → execute → verify. Why this pattern works: Catches errors early: Validation finds problems before changes are applied Machine-verifiable: Scripts provide objective verification Reversible planning: Claude can iterate on the plan without touching originals Clear debugging: Error messages point to specific problems When to use: Batch operations, destructive changes, complex validation rules, high-stakes operations. Implementation tip: Make validation scripts verbose with specific error messages like "Field 'signature_date' not found. Available fields: customer_name, order_total, signature_date_signed" to help Claude fix issues. Package dependencies Skills run in the code execution environment with platform-specific limitations: claude.ai: Can install packages from npm and PyPI and pull from GitHub repositories Anthropic API: Has no network access and no runtime package installation List required packages in your SKILL.md and verify they're available in the code execution tool documentation. Runtime environment Skills run in a code execution environment with filesystem access, bash commands, and code execution capabilities. For the conceptual explanation of this architecture, see The Skills architecture in the overview. How this affects your authoring: How Claude accesses Skills: Metadata pre-loaded: At startup, the name and description from all Skills' YAML frontmatter are loaded into the system prompt Files read on-demand: Claude uses bash Read tools to access SKILL.md and other files from the filesystem when needed Scripts executed efficiently: Utility scripts can be executed via bash without loading their full contents into context. Only the script's output consumes tokens No context penalty for large files: Reference files, data, or documentation don't consume context tokens until actually read File paths matter: Claude navigates your skill directory like a filesystem. Use forward slashes (reference/guide.md), not backslashes Name files descriptively: Use names that indicate content: form_validation_rules.md, not doc2.md Organize for discovery: Structure directories by domain or feature Good: reference/finance.md, reference/sales.md Bad: docs/file1.md, docs/file2.md Bundle comprehensive resources: Include complete API docs, extensive examples, large datasets; no context penalty until accessed Prefer scripts for deterministic operations: Write validate_form.py rather than asking Claude to generate validation code Make execution intent clear: "Run analyze_form.py to extract fields" (execute) "See analyze_form.py for the extraction algorithm" (read as reference) Test file access patterns: Verify Claude can navigate your directory structure by testing with real requests Example: bigquery-skill/ ├── SKILL.md (overview, points to reference files) └── reference/ ├── finance.md (revenue metrics) ├── sales.md (pipeline data) └── product.md (usage analytics) When the user asks about revenue, Claude reads SKILL.md, sees the reference to reference/finance.md, and invokes bash to read just that file. The sales.md and product.md files remain on the filesystem, consuming zero context tokens until needed. This filesystem-based model is what enables progressive disclosure. Claude can navigate and selectively load exactly what each task requires. For complete details on the technical architecture, see How Skills work in the Skills overview. MCP tool references If your Skill uses MCP (Model Context Protocol) tools, always use fully qualified tool names to avoid "tool not found" errors. Format: ServerName:tool_name Example: Use the BigQuery:bigquery_schema tool to retrieve table schemas. Use the GitHub:create_issue tool to create issues. Where: BigQuery and GitHub are MCP server names bigquery_schema and create_issue are the tool names within those servers Without the server prefix, Claude may fail to locate the tool, especially when multiple MCP servers are available. Avoid assuming tools are installed Don't assume packages are available: **Bad example: Assumes installation**: "Use the pdf library to process the file." **Good example: Explicit about dependencies**: "Install required package: `pip install pypdf` Then use it: ```python from pypdf import PdfReader reader = PdfReader("file.pdf") ```" Technical notes YAML frontmatter requirements The SKILL.md frontmatter requires name and description fields with specific validation rules: name: Maximum 64 characters, lowercase letters/numbers/hyphens only, no XML tags, no reserved words See the Skills overview for complete structure details. Token budgets Keep SKILL.md body under 500 lines for optimal performance. If your content exceeds this, split it into separate files using the progressive disclosure patterns described earlier. For architectural details, see the Skills overview. Checklist for effective Skills Before sharing a Skill, verify: Core quality Description is specific and includes key terms Description includes both what the Skill does and when to use it SKILL.md body is under 500 lines Additional details are in separate files (if needed) No time-sensitive information (or in "old patterns" section) Consistent terminology throughout Examples are concrete, not abstract File references are one level deep Progressive disclosure used appropriately Workflows have clear steps Code and scripts Scripts solve problems rather than punt to Claude Error handling is explicit and helpful No "voodoo constants" (all values justified) Required packages listed in instructions and verified as available Scripts have clear documentation No Windows-style paths (all forward slashes) Validation/verification steps for critical operations Feedback loops included for quality-critical tasks Testing At least three evaluations created Tested with Haiku, Sonnet, and Opus Tested with real usage scenarios Team feedback incorporated (if applicable) https://github.com/anthropics/skills --- ## File: docs/blogs.md --- name: context-engineering-blogs doc_type: blog source_url: No --- Some technical blogs that I recently read and find valuable: (Context Engineering 11 min read Jul 2, 2025 TL;DR Agents need context to perform tasks. Context engineering is the art and science of filling the context window with just the right information at each step of an agent’s trajectory. In this post, we break down some common strategies — write, select, compress, and isolate — for context engineering by reviewing various popular agents and papers. We then explain how LangGraph is designed to support them! Also, see our video on context engineering here. General categories of context engineering Context Engineering As Andrej Karpathy puts it, LLMs are like a new kind of operating system. The LLM is like the CPU and its context window is like the RAM, serving as the model’s working memory. Just like RAM, the LLM context window has limited capacity to handle various sources of context. And just as an operating system curates what fits into a CPU’s RAM, we can think about “context engineering” playing a similar role. Karpathy summarizes this well: [Context engineering is the] ”…delicate art and science of filling the context window with just the right information for the next step.” Context types commonly used in LLM applications What are the types of context that we need to manage when building LLM applications? Context engineering as an umbrella that applies across a few different context types: Instructions – prompts, memories, few‑shot examples, tool descriptions, etc Knowledge – facts, memories, etc Tools – feedback from tool calls Context Engineering for Agents This year, interest in agents has grown tremendously as LLMs get better at reasoning and tool calling. Agents interleave LLM invocations and tool calls, often for long-running tasks. Agents interleave LLM calls and tool calls, using tool feedback to decide the next step. Agents interleave LLM calls and tool calls, using tool feedback to decide the next step However, long-running tasks and accumulating feedback from tool calls mean that agents often utilize a large number of tokens. This can cause numerous problems: it can exceed the size of the context window, balloon cost / latency, or degrade agent performance. Drew Breunig nicely outlined a number of specific ways that longer context can cause perform problems, including: Context Poisoning: When a hallucination makes it into the context Context Distraction: When the context overwhelms the training Context Confusion: When superfluous context influences the response Context Clash: When parts of the context disagree Context from tool calls accumulates over multiple agent turns With this in mind, Cognition called out the importance of context engineering: “Context engineering” … is effectively the #1 job of engineers building AI agents. Anthropic also laid it out clearly: Agents often engage in conversations spanning hundreds of turns, requiring careful context management strategies. So, how are people tackling this challenge today? We group common strategies for agent context engineering into four buckets — write, select, compress, and isolate — and give examples of each from review of some popular agent products and papers. We then explain how LangGraph is designed to support them! General categories of context engineering Write Context Writing context means saving it outside the context window to help an agent perform a task. Scratchpads When humans solve tasks, we take notes and remember things for future, related tasks. Agents are also gaining these capabilities! Note-taking via a “scratchpad” is one approach to persist information while an agent is performing a task. The idea is to save information outside of the context window so that it’s available to the agent. Anthropic’s multi-agent researcher illustrates a clear example of this: The LeadResearcher begins by thinking through the approach and saving its plan to Memory to persist the context, since if the context window exceeds 200,000 tokens it will be truncated and it is important to retain the plan. Scratchpads can be implemented in a few different ways. They can be a tool call that simply writes to a file. They can also be a field in a runtime state object that persists during the session. In either case, scratchpads let agents save useful information to help them accomplish a task. Memories Scratchpads help agents solve a task within a given session (or thread), but sometimes agents benefit from remembering things across many sessions! Reflexion introduced the idea of reflection following each agent turn and re-using these self-generated memories. Generative Agents created memories synthesized periodically from collections of past agent feedback. An LLM can be used to update or create memories These concepts made their way into popular products like ChatGPT, Cursor, and Windsurf, which all have mechanisms to auto-generate long-term memories that can persist across sessions based on user-agent interactions. Select Context Selecting context means pulling it into the context window to help an agent perform a task. Scratchpad The mechanism for selecting context from a scratchpad depends upon how the scratchpad is implemented. If it’s a tool, then an agent can simply read it by making a tool call. If it’s part of the agent’s runtime state, then the developer can choose what parts of state to expose to an agent each step. This provides a fine-grained level of control for exposing scratchpad context to the LLM at later turns. Memories If agents have the ability to save memories, they also need the ability to select memories relevant to the task they are performing. This can be useful for a few reasons. Agents might select few-shot examples (episodic memories) for examples of desired behavior, instructions (procedural memories) to steer behavior, or facts (semantic memories) for task-relevant context. One challenge is ensuring that relevant memories are selected. Some popular agents simply use a narrow set of files that are always pulled into context. For example, many code agent use specific files to save instructions (”procedural” memories) or, in some cases, examples (”episodic” memories). Claude Code uses CLAUDE.md. Cursor and Windsurf use rules files. But, if an agent is storing a larger collection of facts and / or relationships (e.g., semantic memories), selection is harder. ChatGPT is a good example of a popular product that stores and selects from a large collection of user-specific memories. Embeddings and / or knowledge graphs for memory indexing are commonly used to assist with selection. Still, memory selection is challenging. At the AIEngineer World’s Fair, Simon Willison shared an example of selection gone wrong: ChatGPT fetched his location from memories and unexpectedly injected it into a requested image. This type of unexpected or undesired memory retrieval can make some users feel like the context window “no longer belongs to them”! Tools Agents use tools, but can become overloaded if they are provided with too many. This is often because the tool descriptions overlap, causing model confusion about which tool to use. One approach is to apply RAG (retrieval augmented generation) to tool descriptions in order to fetch only the most relevant tools for a task. Some recent papers have shown that this improve tool selection accuracy by 3-fold. Knowledge RAG is a rich topic and it can be a central context engineering challenge. Code agents are some of the best examples of RAG in large-scale production. Varun from Windsurf captures some of these challenges well: Indexing code ≠ context retrieval … [We are doing indexing & embedding search … [with] AST parsing code and chunking along semantically meaningful boundaries … embedding search becomes unreliable as a retrieval heuristic as the size of the codebase grows … we must rely on a combination of techniques like grep/file search, knowledge graph based retrieval, and … a re-ranking step where [context] is ranked in order of relevance. Compressing Context Compressing context involves retaining only the tokens required to perform a task. Context Summarization Agent interactions can span hundreds of turns and use token-heavy tool calls. Summarization is one common way to manage these challenges. If you’ve used Claude Code, you’ve seen this in action. Claude Code runs “auto-compact” after you exceed 95% of the context window and it will summarize the full trajectory of user-agent interactions. This type of compression across an agent trajectory can use various strategies such as recursive or hierarchical summarization. A few places where summarization can be applied It can also be useful to add summarization at specific points in an agent’s design. For example, it can be used to post-process certain tool calls (e.g., token-heavy search tools). As a second example, Cognition mentioned summarization at agent-agent boundaries to reduce tokens during knowledge hand-off. Summarization can be a challenge if specific events or decisions need to be captured. Cognition uses a fine-tuned model for this, which underscores how much work can go into this step. Context Trimming Whereas summarization typically uses an LLM to distill the most relevant pieces of context, trimming can often filter or, as Drew Breunig points out, “prune” context. This can use hard-coded heuristics like removing older messages from a list. Drew also mentions Provence, a trained context pruner for Question-Answering. Isolating Context Isolating context involves splitting it up to help an agent perform a task. Multi-agent One of the most popular ways to isolate context is to split it across sub-agents. A motivation for the OpenAI Swarm library was separation of concerns, where a team of agents can handle specific sub-tasks. Each agent has a specific set of tools, instructions, and its own context window. Split context across multiple agents Anthropic’s multi-agent researcher makes a case for this: many agents with isolated contexts outperformed single-agent, largely because each subagent context window can be allocated to a more narrow sub-task. As the blog said: [Subagents operate] in parallel with their own context windows, exploring different aspects of the question simultaneously. Of course, the challenges with multi-agent include token use (e.g., up to 15× more tokens than chat as reported by Anthropic), the need for careful prompt engineering to plan sub-agent work, and coordination of sub-agents. Context Isolation with Environments HuggingFace’s deep researcher shows another interesting example of context isolation. Most agents use tool calling APIs, which return JSON objects (tool arguments) that can be passed to tools (e.g., a search API) to get tool feedback (e.g., search results). HuggingFace uses a CodeAgent, which outputs that contains the desired tool calls. The code then runs in a sandbox. Selected context (e.g., return values) from the tool calls is then passed back to the LLM. Sandboxes can isolate context from the LLM. This allows context to be isolated from the LLM in the environment. Hugging Face noted that this is a great way to isolate token-heavy objects in particular: [Code Agents allow for] a better handling of state … Need to store this image / audio / other for later use? No problem, just assign it as a variable in your state and you [use it later]. State It’s worth calling out that an agent’s runtime state object can also be a great way to isolate context. This can serve the same purpose as sandboxing. A state object can be designed with a schema that has fields that context can be written to. One field of the schema (e.g., messages) can be exposed to the LLM at each turn of the agent, but the schema can isolate information in other fields for more selective use. Context Engineering with LangSmith / LangGraph So, how can you apply these ideas? Before you start, there are two foundational pieces that are helpful. First, ensure that you have a way to look at your data and track token-usage across your agent. This helps inform where best to apply effort context engineering. LangSmith is well-suited for agent tracing / observability, and offers a great way to do this. Second, be sure you have a simple way to test whether context engineering hurts or improve agent performance. LangSmith enables agent evaluation to test the impact of any context engineering effort. Write context LangGraph was designed with both thread-scoped (short-term) and long-term memory. Short-term memory uses checkpointing to persist agent state across all steps of an agent. This is extremely useful as a “scratchpad”, allowing you to write information to state and fetch it at any step in your agent trajectory. LangGraph’s long-term memory lets you to persist context across many sessions with your agent. It is flexible, allowing you to save small sets of files (e.g., a user profile or rules) or larger collections of memories. In addition, LangMem provides a broad set of useful abstractions to aid with LangGraph memory management. Select context Within each node (step) of a LangGraph agent, you can fetch state. This give you fine-grained control over what context you present to the LLM at each agent step. In addition, LangGraph’s long-term memory is accessible within each node and supports various types of retrieval (e.g., fetching files as well as embedding-based retrieval on a memory collection). For an overview of long-term memory, see our Deeplearning.ai course. And for an entry point to memory applied to a specific agent, see our Ambient Agents course. This shows how to use LangGraph memory in a long-running agent that can manage your email and learn from your feedback. Email agent with user feedback and long-term memory For tool selection, the LangGraph Bigtool library is a great way to apply semantic search over tool descriptions. This helps select the most relevant tools for a task when working with a large collection of tools. Finally, we have several tutorials and videos that show how to use various types of RAG with LangGraph. Compressing context Because LangGraph is a low-level orchestration framework, you lay out your agent as a set of nodes, define the logic within each one, and define an state object that is passed between them. This control offers several ways to compress context. One common approach is to use a message list as your agent state and summarize or trim it periodically using a few built-in utilities. However, you can also add logic to post-process tool calls or work phases of your agent in a few different ways. You can add summarization nodes at specific points or also add summarization logic to your tool calling node in order to compress the output of specific tool calls. Isolating context LangGraph is designed around a state object, allowing you to specify a state schema and access state at each agent step. For example, you can store context from tool calls in certain fields in state, isolating them from the LLM until that context is required. In addition to state, LangGraph supports use of sandboxes for context isolation. See this repo for an example LangGraph agent that uses an E2B sandbox for tool calls. See this video for an example of sandboxing using Pyodide where state can be persisted. LangGraph also has a lot of support for building multi-agent architecture, such as the supervisor and swarm libraries. You can see these videos for more detail on using multi-agent with LangGraph. Conclusion Context engineering is becoming a craft that agents builders should aim to master. Here, we covered a few common patterns seen across many popular agents today: Writing context - saving it outside the context window to help an agent perform a task. Selecting context - pulling it into the context window to help an agent perform a task. Compressing context - retaining only the tokens required to perform a task. Isolating context - splitting it up to help an agent perform a task. LangGraph makes it easy to implement each of them and LangSmith provides an easy way to test your agent and track context usage. Together, LangGraph and LangGraph enable a virtuous feedback loop for identifying the best opportunity to apply context engineering, implementing it, testing it, and repeating. --------- Context Engineering in Manus Oct 15, 2025 Lance Martin Why Context Engineering Earlier this week, I had a webinar with Manus co-founder and CSO Yichao “Peak” Ji. You can see the video here, my slides here, and Peak’s slides here. Below are my notes. Anthropic defines agents as systems where LLMs direct their own processes and tool usage, maintaining control over how they accomplish tasks. In short, it’s an LLM calling tools in a loop. Manus is one of the most popular general-purpose consumer agents. The typical Manus task uses 50 tool calls. Without context engineering, these tool call results would accumulate in the LLM context window. As the context window fills, many have observed that LLM performance degrades. For example, Chroma has a great study on context rot and Anthropic has explained how growing context depletes an LLM’s attention budget. So, it’s important to carefully manage what goes into the LLM’s context window when building agents. Karpathy laid this out clearly: Context engineering is the delicate art and science of filling the context window with just the right information for the next step (in an agent’s trajectory) Context Engineering Approaches Each Manus session uses a dedicated cloud-based virtual machine, giving the agent a virtual computer with a filesystem, tools to navigate it, and the ability to execute commands (e.g., provided utilities and standard shell commands) in that sandbox environment. In this sandbox, Manus uses three primary strategies for context engineering, which align with approaches Anthropic covers here and I’ve seen in across many projects: Reduce Context Offload Context Isolate Context Context Reduction Tool calls in Manus have a “full” and “compact” representation. The full version contains the raw content from tool invocation (e.g., a complete search tool result), which is stored in the sandbox (e.g., filesystem). The compact version stores a reference to the full result (e.g., a file path). Manus applies compaction to older (“stale”) tool results. This just means swapping out the full tool result for the compact version. This allows the agent to still fetch the full result if ever needed, but saves tokens by removing “stale” results that the agent has already used to make decisions. Newer tool results remain in full to guide the agent’s next decision. This seems to be a generally useful strategy for context reduction, and I notice that it’s similar to Anthropic’s context editing feature: Context editing automatically clears stale tool calls and results from within the context window when approaching token limits. As your agent executes tasks and accumulates tool results, context editing removes stale content while preserving the conversation flow, effectively extending how long agents can run without manual intervention. When compaction reaches diminishing returns (see figure below), Manus applies summarization to the trajectory. Summaries are generated using full tool results and Manus uses a schema to define the summary fields. This creates a consistent summary object for any agent trajectory. Context Isolation Manus takes a pragmatic approach to multi-agent, avoiding anthropomorphized divisions of labor. While humans organize by role (designer, engineer, project manager) due to cognitive limitations, LLMs don’t necessarily share these same constraints. With this in mind, the primary goal of sub-agents in Manus is to isolate context. For example, if there’s a task to be done, Manus will assign that task to a sub-agent with its own context window. Manus uses multi-agent with a planner that assigns tasks, a knowledge manager that reviews conversations and determines what should be saved in the filesystem, and an executor sub-agent that performs tasks assigned by the planner. Manus initially used a todo.md for task planning, but found that roughly one-third of all actions were spent updating the todo list, wasting valuable tokens. They shifted to a dedicated planner agent that calls executor sub-agents to perform tasks. In a recent podcast, Erik Schluntz (multi-agent research at Anthropic) mentioned that they similarly design multi-agent systems with a planner to assign tasks and use function calling as the communication protocol to initiate sub-agents. A central challenge raised by Erik as well as Walden Yan (Cognition) is context sharing between planner and sub-agents. Manus addresses this in two ways. For simple tasks (e.g., a discrete task where the planner only needs the output of the sub-agent), the planner simply creates instructions and passes them to the sub-agent via the function call. This resembles Claude Code’s task tool. For more complex tasks (e.g., the sub-agent needs to write to files that the planner also uses), the planner shares its full context with the sub-agent. The sub-agent still has its own action space (tools) and instructions, but receives the full context that the planner also has access to. In both cases, the planner defines the sub-agent’s output schema. Sub-agents have a submit results tool to populate this schema before returning results to the planner and Manus uses constrained decoding to ensure output adheres to the defined schema. Context Offloading Tools Definitions We often want agents that can perform a wide range of actions. We can, of course, bind a large collection of tools to the LLM and provide detailed instructions on how to use all of them. But, tool descriptions use valuable tokens and many (often overlapping or ambiguous) tools can cause model confusion. A trend I’m seeing is that agents use a small set of general tools that give the agent access to a computer. For example, with only a Bash tool and a few tools to access a filesystem, an agent can perform a wide range of actions! Manus thinks about this as a layered action space with function/tool calling and its virtual computer sandbox. Peak mentioned that Manus uses a small set (< 20) of atomic functions; this includes things like a Bash tool, tools to manage the filesystem, and a code execution tool. Rather than bloating the function calling layer, Manus offloads most actions to the sandbox layer. Manus can execute many utilities directly in the sandbox with its Bash tool and MCP tools are exposed through a CLI that the agent can also execute using the Bash tool. Claude’s skills feature uses a similar idea:skills are stored in the filesystem, not as bound tools, and Claude only needs a few simple function calls (Bash, file system) to progressively discover and use them. Progressive disclosure is the core design principle that makes Agent Skills flexible and scalable. Like a well-organized manual that starts with a table of contents, then specific chapters, and finally a detailed appendix, skills let Claude load information only as needed … agents with a filesystem and code execution tools don’t need to read the entirety of a skill into their context window when working on a particular task. Tool Results Because Manus has access to a filesystem, it can also offload context (e.g., tool results). As explained above, this is central for context reduction; tool results are offloaded to the filesystem in order to produce the compact version and this is used to prune stale tokens from the agent’s context window. Similar to Claude Code, Manus uses basic utilities (e.g., glob and grep) to search the filesystem without the need for indexing (e.g., vectorstores). Model Choice Rather than committing to a single model, Manus uses task-level routing: it might use Claude for coding, Gemini for multi-modal tasks, or OpenAI for math and reasoning. Broadly, Manus’s approach to model selection is driven by cost considerations, with KV cache efficiency playing a central role. Manus uses caching (e.g., for system instructions, older tool results, etc) to reduce both cost and latency across many agent turns. Peak mentioned that distributed KV cache infrastructure is challenging to implement with open source models, but is well-supported by frontier providers. This caching support can make frontier models cheaper for certain (agent) use-cases in practice. Build with the Bitter Lesson in Mind We closed the discussion talking about the Bitter Lesson. I’ve been interested in its implications for AI engineering. Boris Cherny (creator of Claude Code) mentioned that The Bitter Lesson influenced his decision to keep Claude Code unopinionated, making it easier to adapt to model improvements. Building on constantly improving models means accepting constant change. Peak mentioned that Manus has been refactored five times since their launch in March! In addition, Peak warned that the agent’s harness can limit performance as models advance; this is exactly the challenge called out by the Bitter Lesson. We add structure to improve performance at a point in time, but this structure can limit performance as compute (models) grows. To guard against this, Peak suggested running agent evaluations across varying model strengths. If performance doesn’t improve with stronger models, your harness may be hobbling the agent. This can help test whether your harness is “future proof”. Hyung Won Chung’s (OpenAI/MSL) talk on this topic further emphasizes the need to consistently re-evaluate structure (e.g., your harness / assumptions) as models improve. Add structures needed for the given level of compute and data available. Remove them later, because these shortcuts will bottleneck further improvement. Conclusions Giving agents access to a computer (e.g., filesystem, terminal, utilities) is a common pattern we see across many agents, including Manus. It enables a few context engineering strategies: 1. Offload Context Store tool results externally: Save full tool results to the filesystem (not in context) and access on demand with utilities like glob and grep Push actions to the sandbox: Use a small set of function calls (Bash, filesystem access) that can execute many utilities in the sandbox rather than binding every utility as a tool 2. Reduce Context Compact stale results: Replace older tool results with references (e.g., file paths) as context fills; keep recent results in full to guide the next decision Summarize when needed: Once compaction reaches diminishing returns, apply schema-based summarization to the full trajectory 3. Isolate Context Use sub-agents for discrete tasks: Assign tasks to sub-agents with their own context windows, primarily to isolate context (not to divide labor by role) Share context deliberately: Pass only instructions for simple tasks; pass full context (e.g., trajectory and shared filesystem) for complex tasks where sub-agents need more context A final consideration is to ensure your harness is not limiting performance as models improve (e.g., be “Bitter Lesson-pilled”). Test across model strengths to verify this. Simple, unopinionated designs often adapt better to model improvements. Finally, don’t be afraid to re-build your agent as models improve (Manus refactored 5 times since March)! ----- Context Engineering for AI Agents: Lessons from Building Manus 2025/7/18 --Yichao 'Peak' Ji At the very beginning of the project, my team and I faced a key decision: should we train an end-to-end agentic model using open-source foundations, or build an agent on top of the abilities of frontier models? Back in my first decade in NLP, we didn't have the luxury of that choice. In the distant days of (yes, it's been seven years), models had to be fine-tuned—and evaluated—before they could transfer to a new task. That process often took weeks per iteration, even though the models were tiny compared to today's LLMs. For fast-moving applications, especially pre–PMF, such slow feedback loops are a deal-breaker. That was a bitter lesson from my last startup, where I trained models from scratch for and semantic search. Then came and , and my in-house models became irrelevant overnight. Ironically, those same models marked the beginning of in-context learning—and a whole new path forward. That hard-earned lesson made the choice clear: Manus would bet on context engineering. This allows us to ship improvements in hours instead of weeks, and kept our product orthogonal to the underlying models: If model progress is the rising tide, we want Manus to be the boat, not the pillar stuck to the seabed. Still, context engineering turned out to be anything but straightforward. It's an experimental science—and we've rebuilt our agent framework four times, each time after discovering a better way to shape context. We affectionately refer to this manual process of architecture searching, prompt fiddling, and empirical guesswork as "Stochastic Graduate Descent". It's not elegant, but it works. This post shares the local optima we arrived at through our own "SGD". If you're building your own AI agent, I hope these principles help you converge faster. Design Around the KV-Cache If I had to choose just one metric, I'd argue that the KV-cache hit rate is the single most important metric for a production-stage AI agent. It directly affects both latency and cost. To understand why, let's look at how operates: After receiving a user input, the agent proceeds through a chain of tool uses to complete the task. In each iteration, the model selects an action from a predefined action space based on the current context. That action is then executed in the environment (e.g., Manus's virtual machine sandbox) to produce an observation. The action and observation are appended to the context, forming the input for the next iteration. This loop continues until the task is complete. As you can imagine, the context grows with every step, while the output—usually a structured function call—remains relatively short. This makes the ratio between prefilling and decoding highly skewed in agents compared to chatbots. In Manus, for example, the average input-to-output token ratio is around 100:1. Fortunately, contexts with identical prefixes can take advantage of , which drastically reduces time-to-first-token (TTFT) and inference cost—whether you're using a self-hosted model or calling an inference API. And we're not talking about small savings: with Claude Sonnet, for instance, cached input tokens cost 0.30 USD/MTok, while uncached ones cost 3 USD/MTok—a 10x difference. From a context engineering perspective, improving KV-cache hit rate involves a few key practices: Keep your prompt prefix stable. Due to the nature of LLMs, even a single-token difference can invalidate the cache from that token onward. A common mistake is including a timestamp—especially one precise to the second—at the beginning of the system prompt. Sure, it lets the model tell you the current time, but it also kills your cache hit rate. Make your context append-only. Avoid modifying previous actions or observations. Ensure your serialization is deterministic. Many programming languages and libraries don't guarantee stable key ordering when serializing JSON objects, which can silently break the cache. Mark cache breakpoints explicitly when needed. Some model providers or inference frameworks don't support automatic incremental prefix caching, and instead require manual insertion of cache breakpoints in the context. When assigning these, account for potential cache expiration and at minimum, ensure the breakpoint includes the end of the system prompt. Additionally, if you're self-hosting models using frameworks like , make sure is enabled, and that you're using techniques like session IDs to route requests consistently across distributed workers. Mask, Don't Remove As your agent takes on more capabilities, its action space naturally grows more complex—in plain terms, the number of tools explodes. The recent popularity of only adds fuel to the fire. If you allow user-configurable tools, trust me: someone will inevitably plug hundreds of mysterious tools into your carefully curated action space. As a result, the model is more likely to select the wrong action or take an inefficient path. In short, your heavily armed agent gets dumber. A natural reaction is to design a dynamic action space—perhaps loading tools on demand using something -like. We tried that in Manus too. But our experiments suggest a clear rule: unless absolutely necessary, avoid dynamically adding or removing tools mid-iteration. There are two main reasons for this: In most LLMs, tool definitions live near the front of the context after serialization, typically before or after the system prompt. So any change will invalidate the KV-cache for all subsequent actions and observations. When previous actions and observations still refer to tools that are no longer defined in the current context, the model gets confused. Without , this often leads to schema violations or hallucinated actions. To solve this while still improving action selection, Manus uses a context-aware to manage tool availability. Rather than removing tools, it masks the token logits during decoding to prevent (or enforce) the selection of certain actions based on the current context. In practice, most model providers and inference frameworks support some form of response prefill, which allows you to constrain the action space without modifying the tool definitions. There are generally three modes of function calling (we'll use the from NousResearch as an example): Auto – The model may choose to call a function or not. Implemented by prefilling only the reply prefix: <|im_start|>assistant Required – The model must call a function, but the choice is unconstrained. Implemented by prefilling up to tool call token: <|im_start|>assistant Specified – The model must call a function from a specific subset. Implemented by prefilling up to the beginning of the function name: <|im_start|>assistant{"name": “browser_ Using this, we constrain action selection by masking token logits directly. For example, when the user provides a new input, Manus must reply immediately instead of taking an action. We've also deliberately designed action names with consistent prefixes—e.g., all browser-related tools start with browser_, and command-line tools with shell_. This allows us to easily enforce that the agent only chooses from a certain group of tools at a given state without using stateful logits processors. These designs help ensure that the Manus agent loop remains stable—even under a model-driven architecture. Use the File System as Context Modern frontier LLMs now offer context windows of 128K tokens or more. But in real-world agentic scenarios, that's often not enough, and sometimes even a liability. There are three common pain points: Observations can be huge, especially when agents interact with unstructured data like web pages or PDFs. It's easy to blow past the context limit. Model performance tends to degrade beyond a certain context length, even if the window technically supports it. Long inputs are expensive, even with prefix caching. You're still paying to transmit and prefill every token. To deal with this, many agent systems implement context truncation or compression strategies. But overly aggressive compression inevitably leads to information loss. The problem is fundamental: an agent, by nature, must predict the next action based on all prior state—and you can't reliably predict which observation might become critical ten steps later. From a logical standpoint, any irreversible compression carries risk. That's why we treat the file system as the ultimate context in Manus: unlimited in size, persistent by nature, and directly operable by the agent itself. The model learns to write to and read from files on demand—using the file system not just as storage, but as structured, externalized memory. Our compression strategies are always designed to be restorable. For instance, the content of a web page can be dropped from the context as long as the URL is preserved, and a document's contents can be omitted if its path remains available in the sandbox. This allows Manus to shrink context length without permanently losing information. While developing this feature, I found myself imagining what it would take for a State Space Model (SSM) to work effectively in an agentic setting. Unlike Transformers, SSMs lack full attention and struggle with long-range backward dependencies. But if they could master file-based memory—externalizing long-term state instead of holding it in context—then their speed and efficiency might unlock a new class of agents. Agentic SSMs could be the real successors to . Manipulate Attention Through Recitation If you've worked with Manus, you've probably noticed something curious: when handling complex tasks, it tends to create a todo.md file—and update it step-by-step as the task progresses, checking off completed items. That's not just cute behavior—it's a deliberate mechanism to manipulate attention. A typical task in Manus requires around 50 tool calls on average. That's a long loop—and since Manus relies on LLMs for decision-making, it's vulnerable to drifting off-topic or forgetting earlier goals, especially in long contexts or complicated tasks. By constantly rewriting the todo list, Manus is reciting its objectives into the end of the context. This pushes the global plan into the model's recent attention span, avoiding "lost-in-the-middle" issues and reducing goal misalignment. In effect, it's using natural language to bias its own focus toward the task objective—without needing special architectural changes. Keep the Wrong Stuff In Agents make mistakes. That's not a bug—it's reality. Language models hallucinate, environments return errors, external tools misbehave, and unexpected edge cases show up all the time. In multi-step tasks, failure is not the exception; it's part of the loop. And yet, a common impulse is to hide these errors: clean up the trace, retry the action, or reset the model's state and leave it to the magical "". That feels safer, more controlled. But it comes at a cost: Erasing failure removes evidence. And without evidence, the model can't adapt. In our experience, one of the most effective ways to improve agent behavior is deceptively simple: leave the wrong turns in the context. When the model sees a failed action—and the resulting observation or stack trace—it implicitly updates its internal beliefs. This shifts its prior away from similar actions, reducing the chance of repeating the same mistake. In fact, we believe error recovery is one of the clearest indicators of true agentic behavior. Yet it's still underrepresented in most academic work and public benchmarks, which often focus on task success under ideal conditions. Don't Get Few-Shotted is a common technique for improving LLM outputs. But in agent systems, it can backfire in subtle ways. Language models are excellent mimics; they imitate the pattern of behavior in the context. If your context is full of similar past action-observation pairs, the model will tend to follow that pattern, even when it's no longer optimal. This can be dangerous in tasks that involve repetitive decisions or actions. For example, when using Manus to help review a batch of 20 resumes, the agent often falls into a rhythm—repeating similar actions simply because that's what it sees in the context. This leads to drift, overgeneralization, or sometimes hallucination. The fix is to increase diversity. Manus introduces small amounts of structured variation in actions and observations—different serialization templates, alternate phrasing, minor noise in order or formatting. This controlled randomness helps break the pattern and tweaks the model's attention. In other words, don't few-shot yourself into a rut. The more uniform your context, the more brittle your agent becomes. Conclusion Context engineering is still an emerging science—but for agent systems, it's already essential. Models may be getting stronger, faster, and cheaper, but no amount of raw capability replaces the need for memory, environment, and feedback. How you shape the context ultimately defines how your agent behaves: how fast it runs, how well it recovers, and how far it scales. At Manus, we've learned these lessons through repeated rewrites, dead ends, and real-world testing across millions of users. None of what we've shared here is universal truth—but these are the patterns that worked for us. If they help you avoid even one painful iteration, then this post did its job. The agentic future will be built one context at a time. Engineer them well. ------ Wide Research: Beyond the Context Window The promise of AI-driven research has always been compelling: delegate the tedious work of information gathering and synthesis to an intelligent system, freeing up human cognition for higher-order analysis and decision-making. Yet, anyone who has pushed these systems on non-trivial use cases has run into a frustrating reality: by the eighth or ninth item in a multi-subject research task, the AI starts fabricating. Not just simplifying. Not just summarizing more concisely. Fabricating. This isn't a prompt engineering problem. It's not a model capability problem. It is an architectural constraint that has quietly limited the utility of AI research tools since their inception. And it's the constraint that Wide Research is designed to overcome. The Context Window: A Fundamental Bottleneck Every large language model operates within a context window, a finite memory buffer that limits the amount of information the model can actively process at any given moment. Modern models have pushed this boundary impressively: from 4K tokens to 32K, 128K, and even 1M tokens in recent versions. Yet the problem persists. When you ask an AI to research multiple entities-say, fifty companies, thirty research papers, or twenty competing products-the context window fills up rapidly. It's not just the raw information about each entity, but also: The original task specification and requirements The structural template for consistent output formatting Intermediate reasoning and analysis for each item Cross-referencing and comparative notes The cumulative context of all preceding items By the time the model reaches the eighth or ninth item, the context window is under immense strain. The model faces an impossible choice: fail explicitly, or start cutting corners. It always chooses the latter. The Fabrication Threshold Here's what happens in practice: Items 1-5: The model performs genuine research. It retrieves information, cross-references sources, and produces detailed, accurate analysis. Items 6-8: The quality begins to subtly degrade. Descriptions become slightly more generic. The model starts relying more on prior patterns than fresh research. Items 9+: The model enters fabrication mode. Unable to maintain the cognitive load of thorough research while managing an overflowing context, it begins generating plausible-sounding content based on statistical patterns, not actual investigation. These fabrications are sophisticated. They sound authoritative. They follow the established format perfectly. They are often grammatically flawless and stylistically consistent with the earlier, legitimate entries. They are also frequently wrong. A competitor analysis might attribute features to companies that don't offer them. A literature review might cite papers with fabricated findings. A product comparison might invent pricing tiers or specifications. The insidious part is that these fabrications are difficult to detect without manual verification—which defeats the entire purpose of automated research. Why Bigger Context Windows Can't Fix This The intuitive response is to simply expand the context window. If 32K tokens aren't enough, use 128K. If that's not enough, push to 200K or beyond. This approach misunderstands the problem. First, context decay is not binary. A model does not maintain perfect recall across its entire context window. Studies have shown that retrieval accuracy degrades with distance from the current position—the "lost in the middle" phenomenon. Information at the beginning and end of the context is recalled more reliably than information in the middle. Second, the processing cost grows disproportionately. The cost to process a 400K token context isn't just double the cost of 200K—it increases exponentially in both time and computing resources. This makes massive-context processing economically impractical for many use cases. Third, the problem is cognitive load. Even with an infinite context, asking a single model to maintain consistent quality across dozens of independent research tasks creates a cognitive bottleneck. The model must constantly switch context between items, maintain a comparative framework, and ensure stylistic consistency—all while performing the core research task. Fourth, context length pressure. The model’s “patience” is, to some extent, determined by the length distribution of samples in its training data. However, the post-training data mixture of current language models is still dominated by relatively short trajectories designed for chatbot-style interactions. As a result, when the length of an assistant message’s content exceeds a certain threshold, the model naturally experiences a kind of context length pressure, prompting it to hasten toward summarizing or to resort to incomplete expression forms such as bullet points. The context window is a constraint, yes. But it's a symptom of a deeper architectural limitation: the single-processor, sequential paradigm. The Architectural Shift: Parallel Processing Wide Research represents a fundamental rethinking of how an AI system should approach large-scale research tasks. Instead of asking one processor to handle n items sequentially, we deploy n parallel sub-agents to process n items simultaneously. The Wide Research Architecture When you launch a Wide Research task, the system operates as follows: 1. Intelligent Decomposition The main controller analyzes your request and breaks it down into independent, parallelizable sub-tasks. This involves understanding the task structure, identifying dependencies, and creating coherent sub-specifications. 2. Sub-agent Delegation For each sub-task, the system spins up a dedicated sub-agent. Crucially, these are not lightweight processes—they are full-featured Manus instances, each with: A complete virtual machine environment Access to the full tool library (search, browsing, code execution, file handling) An independent internet connection A fresh, empty context window 3. Parallel Execution All sub-agents execute simultaneously. Each one focuses exclusively on its assigned item, performing the same depth of research and analysis it would for a single-item task. 4. Centralized Coordination The main controller maintains oversight, collecting results as the sub-agents complete their jobs. Importantly, the sub-agents do not communicate with each other, all coordination flows through the main controller. This prevents context pollution and maintains independence. 5. Synthesis and Integration Once all sub-agents have reported back, the main controller synthesizes the results into a single, coherent, and comprehensive report. This synthesis step leverages the full context capacity of the main controller, as it is not burdened with the original research effort. Why This Changes Everything Consistent Quality at Scale Every item gets the same treatment. The 50th item is researched just as thoroughly as the first. There is no degradation curve, no fabrication threshold, and no quality cliff. True Horizontal Scalability Need to analyze 10 items? The system deploys 10 sub-agents. Need to analyze 500? It deploys 500. The architecture scales linearly with the size of the task, not exponentially like context-based approaches. Significant Speed-Up Because the sub-agents operate in parallel, the real-world time required to analyze 50 items is roughly the same as the time to analyze 5. The bottleneck shifts from sequential processing time to synthesis time—a much smaller component of the overall task. Reduced Hallucination Rate Each sub-agent operates within its cognitive comfort zone. With a fresh context and a single, focused task, there is no pressure to fabricate. The sub-agent can perform genuine research, verify facts, and maintain accuracy. Independence and Reliability Because the sub-agents do not share context, an error or hallucination in one sub-agent's job does not propagate to the others. Each analysis stands on its own, reducing systemic risk. Beyond the Single-Processor Paradigm Wide Research is more than a feature—it represents a fundamental shift away from the single-processor paradigm and toward an orchestrated, parallel architecture. The future of AI systems lies not in ever-larger context windows, but in intelligent task decomposition and parallel execution. We are moving from the era of the "AI assistant" to the era of the "AI workforce." When to use Wide Research: Any task involving multiple, similar items that require consistent analysiscompetitive research, literature reviews, bulk processing, multi-asset generation. When not to use: Deeply sequential tasks where each step heavily depends on the prior result, or small tasks (fewer than 10 items) where single-processor handling is more cost-effective. -------- How we built our multi-agent research system Published Jun 13, 2025 Our Research feature uses multiple Claude agents to explore complex topics more effectively. We share the engineering challenges and the lessons we learned from building this system. Claude now has Research capabilities that allow it to search across the web, Google Workspace, and any integrations to accomplish complex tasks. The journey of this multi-agent system from prototype to production taught us critical lessons about system architecture, tool design, and prompt engineering. A multi-agent system consists of multiple agents (LLMs autonomously using tools in a loop) working together. Our Research feature involves an agent that plans a research process based on user queries, and then uses tools to create parallel agents that search for information simultaneously. Systems with multiple agents introduce new challenges in agent coordination, evaluation, and reliability. This post breaks down the principles that worked for us—we hope you'll find them useful to apply when building your own multi-agent systems. Benefits of a multi-agent system Research work involves open-ended problems where it’s very difficult to predict the required steps in advance. You can’t hardcode a fixed path for exploring complex topics, as the process is inherently dynamic and path-dependent. When people conduct research, they tend to continuously update their approach based on discoveries, following leads that emerge during investigation. This unpredictability makes AI agents particularly well-suited for research tasks. Research demands the flexibility to pivot or explore tangential connections as the investigation unfolds. The model must operate autonomously for many turns, making decisions about which directions to pursue based on intermediate findings. A linear, one-shot pipeline cannot handle these tasks. The essence of search is compression: distilling insights from a vast corpus. Subagents facilitate compression by operating in parallel with their own context windows, exploring different aspects of the question simultaneously before condensing the most important tokens for the lead research agent. Each subagent also provides separation of concerns—distinct tools, prompts, and exploration trajectories—which reduces path dependency and enables thorough, independent investigations. Once intelligence reaches a threshold, multi-agent systems become a vital way to scale performance. For instance, although individual humans have become more intelligent in the last 100,000 years, human societies have become exponentially more capable in the information age because of our collective intelligence and ability to coordinate. Even generally-intelligent agents face limits when operating as individuals; groups of agents can accomplish far more. Our internal evaluations show that multi-agent research systems excel especially for breadth-first queries that involve pursuing multiple independent directions simultaneously. We found that a multi-agent system with Claude Opus 4 as the lead agent and Claude Sonnet 4 subagents outperformed single-agent Claude Opus 4 by 90.2% on our internal research eval. For example, when asked to identify all the board members of the companies in the Information Technology S&P 500, the multi-agent system found the correct answers by decomposing this into tasks for subagents, while the single agent system failed to find the answer with slow, sequential searches. Multi-agent systems work mainly because they help spend enough tokens to solve the problem. In our analysis, three factors explained 95% of the performance variance in the BrowseComp evaluation (which tests the ability of browsing agents to locate hard-to-find information). We found that token usage by itself explains 80% of the variance, with the number of tool calls and the model choice as the two other explanatory factors. This finding validates our architecture that distributes work across agents with separate context windows to add more capacity for parallel reasoning. The latest Claude models act as large efficiency multipliers on token use, as upgrading to Claude Sonnet 4 is a larger performance gain than doubling the token budget on Claude Sonnet 3.7. Multi-agent architectures effectively scale token usage for tasks that exceed the limits of single agents. There is a downside: in practice, these architectures burn through tokens fast. In our data, agents typically use about 4× more tokens than chat interactions, and multi-agent systems use about 15× more tokens than chats. For economic viability, multi-agent systems require tasks where the value of the task is high enough to pay for the increased performance. Further, some domains that require all agents to share the same context or involve many dependencies between agents are not a good fit for multi-agent systems today. For instance, most coding tasks involve fewer truly parallelizable tasks than research, and LLM agents are not yet great at coordinating and delegating to other agents in real time. We’ve found that multi-agent systems excel at valuable tasks that involve heavy parallelization, information that exceeds single context windows, and interfacing with numerous complex tools. Architecture overview for Research Our Research system uses a multi-agent architecture with an orchestrator-worker pattern, where a lead agent coordinates the process while delegating to specialized subagents that operate in parallel. The multi-agent architecture in action: user queries flow through a lead agent that creates specialized subagents to search for different aspects in parallel. When a user submits a query, the lead agent analyzes it, develops a strategy, and spawns subagents to explore different aspects simultaneously. As shown in the diagram above, the subagents act as intelligent filters by iteratively using search tools to gather information, in this case on AI agent companies in 2025, and then returning a list of companies to the lead agent so it can compile a final answer. Traditional approaches using Retrieval Augmented Generation (RAG) use static retrieval. That is, they fetch some set of chunks that are most similar to an input query and use these chunks to generate a response. In contrast, our architecture uses a multi-step search that dynamically finds relevant information, adapts to new findings, and analyzes results to formulate high-quality answers. Process diagram showing the complete workflow of our multi-agent Research system. When a user submits a query, the system creates a LeadResearcher agent that enters an iterative research process. The LeadResearcher begins by thinking through the approach and saving its plan to Memory to persist the context, since if the context window exceeds 200,000 tokens it will be truncated and it is important to retain the plan. It then creates specialized Subagents (two are shown here, but it can be any number) with specific research tasks. Each Subagent independently performs web searches, evaluates tool results using interleaved thinking, and returns findings to the LeadResearcher. The LeadResearcher synthesizes these results and decides whether more research is needed—if so, it can create additional subagents or refine its strategy. Once sufficient information is gathered, the system exits the research loop and passes all findings to a CitationAgent, which processes the documents and research report to identify specific locations for citations. This ensures all claims are properly attributed to their sources. The final research results, complete with citations, are then returned to the user. Prompt engineering and evaluations for research agents Multi-agent systems have key differences from single-agent systems, including a rapid growth in coordination complexity. Early agents made errors like spawning 50 subagents for simple queries, scouring the web endlessly for nonexistent sources, and distracting each other with excessive updates. Since each agent is steered by a prompt, prompt engineering was our primary lever for improving these behaviors. Below are some principles we learned for prompting agents: Think like your agents. To iterate on prompts, you must understand their effects. To help us do this, we built simulations using our Console with the exact prompts and tools from our system, then watched agents work step-by-step. This immediately revealed failure modes: agents continuing when they already had sufficient results, using overly verbose search queries, or selecting incorrect tools. Effective prompting relies on developing an accurate mental model of the agent, which can make the most impactful changes obvious. Teach the orchestrator how to delegate. In our system, the lead agent decomposes queries into subtasks and describes them to subagents. Each subagent needs an objective, an output format, guidance on the tools and sources to use, and clear task boundaries. Without detailed task descriptions, agents duplicate work, leave gaps, or fail to find necessary information. We started by allowing the lead agent to give simple, short instructions like 'research the semiconductor shortage,' but found these instructions often were vague enough that subagents misinterpreted the task or performed the exact same searches as other agents. For instance, one subagent explored the 2021 automotive chip crisis while 2 others duplicated work investigating current 2025 supply chains, without an effective division of labor. Scale effort to query complexity. Agents struggle to judge appropriate effort for different tasks, so we embedded scaling rules in the prompts. Simple fact-finding requires just 1 agent with 3-10 tool calls, direct comparisons might need 2-4 subagents with 10-15 calls each, and complex research might use more than 10 subagents with clearly divided responsibilities. These explicit guidelines help the lead agent allocate resources efficiently and prevent overinvestment in simple queries, which was a common failure mode in our early versions. Tool design and selection are critical. Agent-tool interfaces are as critical as human-computer interfaces. Using the right tool is efficient—often, it’s strictly necessary. For instance, an agent searching the web for context that only exists in Slack is doomed from the start. With MCP servers that give the model access to external tools, this problem compounds, as agents encounter unseen tools with descriptions of wildly varying quality. We gave our agents explicit heuristics: for example, examine all available tools first, match tool usage to user intent, search the web for broad external exploration, or prefer specialized tools over generic ones. Bad tool descriptions can send agents down completely wrong paths, so each tool needs a distinct purpose and a clear description. Let agents improve themselves. We found that the Claude 4 models can be excellent prompt engineers. When given a prompt and a failure mode, they are able to diagnose why the agent is failing and suggest improvements. We even created a tool-testing agent—when given a flawed MCP tool, it attempts to use the tool and then rewrites the tool description to avoid failures. By testing the tool dozens of times, this agent found key nuances and bugs. This process for improving tool ergonomics resulted in a 40% decrease in task completion time for future agents using the new description, because they were able to avoid most mistakes. Start wide, then narrow down. Search strategy should mirror expert human research: explore the landscape before drilling into specifics. Agents often default to overly long, specific queries that return few results. We counteracted this tendency by prompting agents to start with short, broad queries, evaluate what’s available, then progressively narrow focus. Guide the thinking process. Extended thinking mode, which leads Claude to output additional tokens in a visible thinking process, can serve as a controllable scratchpad. The lead agent uses thinking to plan its approach, assessing which tools fit the task, determining query complexity and subagent count, and defining each subagent’s role. Our testing showed that extended thinking improved instruction-following, reasoning, and efficiency. Subagents also plan, then use interleaved thinking after tool results to evaluate quality, identify gaps, and refine their next query. This makes subagents more effective in adapting to any task. Parallel tool calling transforms speed and performance. Complex research tasks naturally involve exploring many sources. Our early agents executed sequential searches, which was painfully slow. For speed, we introduced two kinds of parallelization: (1) the lead agent spins up 3-5 subagents in parallel rather than serially; (2) the subagents use 3+ tools in parallel. These changes cut research time by up to 90% for complex queries, allowing Research to do more work in minutes instead of hours while covering more information than other systems. Our prompting strategy focuses on instilling good heuristics rather than rigid rules. We studied how skilled humans approach research tasks and encoded these strategies in our prompts—strategies like decomposing difficult questions into smaller tasks, carefully evaluating the quality of sources, adjusting search approaches based on new information, and recognizing when to focus on depth (investigating one topic in detail) vs. breadth (exploring many topics in parallel). We also proactively mitigated unintended side effects by setting explicit guardrails to prevent the agents from spiraling out of control. Finally, we focused on a fast iteration loop with observability and test cases. Effective evaluation of agents Good evaluations are essential for building reliable AI applications, and agents are no different. However, evaluating multi-agent systems presents unique challenges. Traditional evaluations often assume that the AI follows the same steps each time: given input X, the system should follow path Y to produce output Z. But multi-agent systems don't work this way. Even with identical starting points, agents might take completely different valid paths to reach their goal. One agent might search three sources while another searches ten, or they might use different tools to find the same answer. Because we don’t always know what the right steps are, we usually can't just check if agents followed the “correct” steps we prescribed in advance. Instead, we need flexible evaluation methods that judge whether agents achieved the right outcomes while also following a reasonable process. Start evaluating immediately with small samples. In early agent development, changes tend to have dramatic impacts because there is abundant low-hanging fruit. A prompt tweak might boost success rates from 30% to 80%. With effect sizes this large, you can spot changes with just a few test cases. We started with a set of about 20 queries representing real usage patterns. Testing these queries often allowed us to clearly see the impact of changes. We often hear that AI developer teams delay creating evals because they believe that only large evals with hundreds of test cases are useful. However, it’s best to start with small-scale testing right away with a few examples, rather than delaying until you can build more thorough evals. LLM-as-judge evaluation scales when done well. Research outputs are difficult to evaluate programmatically, since they are free-form text and rarely have a single correct answer. LLMs are a natural fit for grading outputs. We used an LLM judge that evaluated each output against criteria in a rubric: factual accuracy (do claims match sources?), citation accuracy (do the cited sources match the claims?), completeness (are all requested aspects covered?), source quality (did it use primary sources over lower-quality secondary sources?), and tool efficiency (did it use the right tools a reasonable number of times?). We experimented with multiple judges to evaluate each component, but found that a single LLM call with a single prompt outputting scores from 0.0-1.0 and a pass-fail grade was the most consistent and aligned with human judgements. This method was especially effective when the eval test cases did have a clear answer, and we could use the LLM judge to simply check if the answer was correct (i.e. did it accurately list the pharma companies with the top 3 largest R&D budgets?). Using an LLM as a judge allowed us to scalably evaluate hundreds of outputs. Human evaluation catches what automation misses. People testing agents find edge cases that evals miss. These include hallucinated answers on unusual queries, system failures, or subtle source selection biases. In our case, human testers noticed that our early agents consistently chose SEO-optimized content farms over authoritative but less highly-ranked sources like academic PDFs or personal blogs. Adding source quality heuristics to our prompts helped resolve this issue. Even in a world of automated evaluations, manual testing remains essential. Multi-agent systems have emergent behaviors, which arise without specific programming. For instance, small changes to the lead agent can unpredictably change how subagents behave. Success requires understanding interaction patterns, not just individual agent behavior. Therefore, the best prompts for these agents are not just strict instructions, but frameworks for collaboration that define the division of labor, problem-solving approaches, and effort budgets. Getting this right relies on careful prompting and tool design, solid heuristics, observability, and tight feedback loops. See the open-source prompts in our Cookbook for example prompts from our system. Production reliability and engineering challenges In traditional software, a bug might break a feature, degrade performance, or cause outages. In agentic systems, minor changes cascade into large behavioral changes, which makes it remarkably difficult to write code for complex agents that must maintain state in a long-running process. Agents are stateful and errors compound. Agents can run for long periods of time, maintaining state across many tool calls. This means we need to durably execute code and handle errors along the way. Without effective mitigations, minor system failures can be catastrophic for agents. When errors occur, we can't just restart from the beginning: restarts are expensive and frustrating for users. Instead, we built systems that can resume from where the agent was when the errors occurred. We also use the model’s intelligence to handle issues gracefully: for instance, letting the agent know when a tool is failing and letting it adapt works surprisingly well. We combine the adaptability of AI agents built on Claude with deterministic safeguards like retry logic and regular checkpoints. Debugging benefits from new approaches. Agents make dynamic decisions and are non-deterministic between runs, even with identical prompts. This makes debugging harder. For instance, users would report agents “not finding obvious information,” but we couldn't see why. Were the agents using bad search queries? Choosing poor sources? Hitting tool failures? Adding full production tracing let us diagnose why agents failed and fix issues systematically. Beyond standard observability, we monitor agent decision patterns and interaction structures—all without monitoring the contents of individual conversations, to maintain user privacy. This high-level observability helped us diagnose root causes, discover unexpected behaviors, and fix common failures. Deployment needs careful coordination. Agent systems are highly stateful webs of prompts, tools, and execution logic that run almost continuously. This means that whenever we deploy updates, agents might be anywhere in their process. We therefore need to prevent our well-meaning code changes from breaking existing agents. We can’t update every agent to the new version at the same time. Instead, we use rainbow deployments to avoid disrupting running agents, by gradually shifting traffic from old to new versions while keeping both running simultaneously. Synchronous execution creates bottlenecks. Currently, our lead agents execute subagents synchronously, waiting for each set of subagents to complete before proceeding. This simplifies coordination, but creates bottlenecks in the information flow between agents. For instance, the lead agent can’t steer subagents, subagents can’t coordinate, and the entire system can be blocked while waiting for a single subagent to finish searching. Asynchronous execution would enable additional parallelism: agents working concurrently and creating new subagents when needed. But this asynchronicity adds challenges in result coordination, state consistency, and error propagation across the subagents. As models can handle longer and more complex research tasks, we expect the performance gains will justify the complexity. Conclusion When building AI agents, the last mile often becomes most of the journey. Codebases that work on developer machines require significant engineering to become reliable production systems. The compound nature of errors in agentic systems means that minor issues for traditional software can derail agents entirely. One step failing can cause agents to explore entirely different trajectories, leading to unpredictable outcomes. For all the reasons described in this post, the gap between prototype and production is often wider than anticipated. Despite these challenges, multi-agent systems have proven valuable for open-ended research tasks. Users have said that Claude helped them find business opportunities they hadn’t considered, navigate complex healthcare options, resolve thorny technical bugs, and save up to days of work by uncovering research connections they wouldn't have found alone. Multi-agent research systems can operate reliably at scale with careful engineering, comprehensive testing, detail-oriented prompt and tool design, robust operational practices, and tight collaboration between research, product, and engineering teams who have a strong understanding of current agent capabilities. We're already seeing these systems transform how people solve complex problems. A Clio embedding plot showing the most common ways people are using the Research feature today. The top use case categories are developing software systems across specialized domains (10%), develop and optimize professional and technical content (8%), develop business growth and revenue generation strategies (8%), assist with academic research and educational material development (7%), and research and verify information about people, places, or organizations (5%). Acknowledgements Written by Jeremy Hadfield, Barry Zhang, Kenneth Lien, Florian Scholz, Jeremy Fox, and Daniel Ford. This work reflects the collective efforts of several teams across Anthropic who made the Research feature possible. Special thanks go to the Anthropic apps engineering team, whose dedication brought this complex multi-agent system to production. We're also grateful to our early users for their excellent feedback. Appendix Below are some additional miscellaneous tips for multi-agent systems. End-state evaluation of agents that mutate state over many turns. Evaluating agents that modify persistent state across multi-turn conversations presents unique challenges. Unlike read-only research tasks, each action can change the environment for subsequent steps, creating dependencies that traditional evaluation methods struggle to handle. We found success focusing on end-state evaluation rather than turn-by-turn analysis. Instead of judging whether the agent followed a specific process, evaluate whether it achieved the correct final state. This approach acknowledges that agents may find alternative paths to the same goal while still ensuring they deliver the intended outcome. For complex workflows, break evaluation into discrete checkpoints where specific state changes should have occurred, rather than attempting to validate every intermediate step. Long-horizon conversation management. Production agents often engage in conversations spanning hundreds of turns, requiring careful context management strategies. As conversations extend, standard context windows become insufficient, necessitating intelligent compression and memory mechanisms. We implemented patterns where agents summarize completed work phases and store essential information in external memory before proceeding to new tasks. When context limits approach, agents can spawn fresh subagents with clean contexts while maintaining continuity through careful handoffs. Further, they can retrieve stored context like the research plan from their memory rather than losing previous work when reaching the context limit. This distributed approach prevents context overflow while preserving conversation coherence across extended interactions. Subagent output to a filesystem to minimize the ‘game of telephone.’ Direct subagent outputs can bypass the main coordinator for certain types of results, improving both fidelity and performance. Rather than requiring subagents to communicate everything through the lead agent, implement artifact systems where specialized agents can create outputs that persist independently. Subagents call tools to store their work in external systems, then pass lightweight references back to the coordinator. This prevents information loss during multi-stage processing and reduces token overhead from copying large outputs through conversation history. The pattern works particularly well for structured outputs like code, reports, or data visualizations where the subagent's specialized prompt produces better results than filtering through a general coordinator. ------- riting effective tools for agents — with agents Published Sep 11, 2025 Agents are only as effective as the tools we give them. We share how to write high-quality tools and evaluations, and how you can boost performance by using Claude to optimize its tools for itself. The Model Context Protocol (MCP) can empower LLM agents with potentially hundreds of tools to solve real-world tasks. But how do we make those tools maximally effective? In this post, we describe our most effective techniques for improving performance in a variety of agentic AI systems1. We begin by covering how you can: Build and test prototypes of your tools Create and run comprehensive evaluations of your tools with agents Collaborate with agents like Claude Code to automatically increase the performance of your tools We conclude with key principles for writing high-quality tools we’ve identified along the way: Choosing the right tools to implement (and not to implement) Namespacing tools to define clear boundaries in functionality Returning meaningful context from tools back to agents Optimizing tool responses for token efficiency Prompt-engineering tool descriptions and specs This is an image depicting how an engineer might use Claude Code to evaluate the efficacy of agentic tools. Building an evaluation allows you to systematically measure the performance of your tools. You can use Claude Code to automatically optimize your tools against this evaluation. What is a tool? In computing, deterministic systems produce the same output every time given identical inputs, while non-deterministic systems—like agents—can generate varied responses even with the same starting conditions. When we traditionally write software, we’re establishing a contract between deterministic systems. For instance, a function call like getWeather(“NYC”) will always fetch the weather in New York City in the exact same manner every time it is called. Tools are a new kind of software which reflects a contract between deterministic systems and non-deterministic agents. When a user asks "Should I bring an umbrella today?,” an agent might call the weather tool, answer from general knowledge, or even ask a clarifying question about location first. Occasionally, an agent might hallucinate or even fail to grasp how to use a tool. This means fundamentally rethinking our approach when writing software for agents: instead of writing tools and MCP servers the way we’d write functions and APIs for other developers or systems, we need to design them for agents. Our goal is to increase the surface area over which agents can be effective in solving a wide range of tasks by using tools to pursue a variety of successful strategies. Fortunately, in our experience, the tools that are most “ergonomic” for agents also end up being surprisingly intuitive to grasp as humans. How to write tools In this section, we describe how you can collaborate with agents both to write and to improve the tools you give them. Start by standing up a quick prototype of your tools and testing them locally. Next, run a comprehensive evaluation to measure subsequent changes. Working alongside agents, you can repeat the process of evaluating and improving your tools until your agents achieve strong performance on real-world tasks. Building a prototype It can be difficult to anticipate which tools agents will find ergonomic and which tools they won’t without getting hands-on yourself. Start by standing up a quick prototype of your tools. If you’re using Claude Code to write your tools (potentially in one-shot), it helps to give Claude documentation for any software libraries, APIs, or SDKs (including potentially the MCP SDK) your tools will rely on. LLM-friendly documentation can commonly be found in flat llms.txt files on official documentation sites (here’s our API’s). Wrapping your tools in a local MCP server or Desktop extension (DXT) will allow you to connect and test your tools in Claude Code or the Claude Desktop app. To connect your local MCP server to Claude Code, run claude mcp add [args...]. To connect your local MCP server or DXT to the Claude Desktop app, navigate to Settings > Developer or Settings > Extensions, respectively. Tools can also be passed directly into Anthropic API calls for programmatic testing. Test the tools yourself to identify any rough edges. Collect feedback from your users to build an intuition around the use-cases and prompts you expect your tools to enable. Running an evaluation Next, you need to measure how well Claude uses your tools by running an evaluation. Start by generating lots of evaluation tasks, grounded in real world uses. We recommend collaborating with an agent to help analyze your results and determine how to improve your tools. See this process end-to-end in our tool evaluation cookbook. This graph measures the test set accuracy of human-written vs. Claude-optimized Slack MCP servers. Held-out test set performance of our internal Slack tools Generating evaluation tasks With your early prototype, Claude Code can quickly explore your tools and create dozens of prompt and response pairs. Prompts should be inspired by real-world uses and be based on realistic data sources and services (for example, internal knowledge bases and microservices). We recommend you avoid overly simplistic or superficial “sandbox” environments that don’t stress-test your tools with sufficient complexity. Strong evaluation tasks might require multiple tool calls—potentially dozens. Here are some examples of strong tasks: Schedule a meeting with Jane next week to discuss our latest Acme Corp project. Attach the notes from our last project planning meeting and reserve a conference room. Customer ID 9182 reported that they were charged three times for a single purchase attempt. Find all relevant log entries and determine if any other customers were affected by the same issue. Customer Sarah Chen just submitted a cancellation request. Prepare a retention offer. Determine: (1) why they're leaving, (2) what retention offer would be most compelling, and (3) any risk factors we should be aware of before making an offer. And here are some weaker tasks: Schedule a meeting with jane@acme.corp next week. Search the payment logs for purchase_complete and customer_id=9182. Find the cancellation request by Customer ID 45892. Each evaluation prompt should be paired with a verifiable response or outcome. Your verifier can be as simple as an exact string comparison between ground truth and sampled responses, or as advanced as enlisting Claude to judge the response. Avoid overly strict verifiers that reject correct responses due to spurious differences like formatting, punctuation, or valid alternative phrasings. For each prompt-response pair, you can optionally also specify the tools you expect an agent to call in solving the task, to measure whether or not agents are successful in grasping each tool’s purpose during evaluation. However, because there might be multiple valid paths to solving tasks correctly, try to avoid overspecifying or overfitting to strategies. Running the evaluation We recommend running your evaluation programmatically with direct LLM API calls. Use simple agentic loops (while-loops wrapping alternating LLM API and tool calls): one loop for each evaluation task. Each evaluation agent should be given a single task prompt and your tools. In your evaluation agents’ system prompts, we recommend instructing agents to output not just structured response blocks (for verification), but also reasoning and feedback blocks. Instructing agents to output these before tool call and response blocks may increase LLMs’ effective intelligence by triggering chain-of-thought (CoT) behaviors. If you’re running your evaluation with Claude, you can turn on interleaved thinking for similar functionality “off-the-shelf”. This will help you probe why agents do or don’t call certain tools and highlight specific areas of improvement in tool descriptions and specs. As well as top-level accuracy, we recommend collecting other metrics like the total runtime of individual tool calls and tasks, the total number of tool calls, the total token consumption, and tool errors. Tracking tool calls can help reveal common workflows that agents pursue and offer some opportunities for tools to consolidate. This graph measures the test set accuracy of human-written vs. Claude-optimized Asana MCP servers. Held-out test set performance of our internal Asana tools Analyzing results Agents are your helpful partners in spotting issues and providing feedback on everything from contradictory tool descriptions to inefficient tool implementations and confusing tool schemas. However, keep in mind that what agents omit in their feedback and responses can often be more important than what they include. LLMs don’t always say what they mean. Observe where your agents get stumped or confused. Read through your evaluation agents’ reasoning and feedback (or CoT) to identify rough edges. Review the raw transcripts (including tool calls and tool responses) to catch any behavior not explicitly described in the agent’s CoT. Read between the lines; remember that your evaluation agents don’t necessarily know the correct answers and strategies. Analyze your tool calling metrics. Lots of redundant tool calls might suggest some rightsizing of pagination or token limit parameters is warranted; lots of tool errors for invalid parameters might suggest tools could use clearer descriptions or better examples. When we launched Claude’s web search tool, we identified that Claude was needlessly appending 2025 to the tool’s query parameter, biasing search results and degrading performance (we steered Claude in the right direction by improving the tool description). Collaborating with agents You can even let agents analyze your results and improve your tools for you. Simply concatenate the transcripts from your evaluation agents and paste them into Claude Code. Claude is an expert at analyzing transcripts and refactoring lots of tools all at once—for example, to ensure tool implementations and descriptions remain self-consistent when new changes are made. In fact, most of the advice in this post came from repeatedly optimizing our internal tool implementations with Claude Code. Our evaluations were created on top of our internal workspace, mirroring the complexity of our internal workflows, including real projects, documents, and messages. We relied on held-out test sets to ensure we did not overfit to our “training” evaluations. These test sets revealed that we could extract additional performance improvements even beyond what we achieved with "expert" tool implementations—whether those tools were manually written by our researchers or generated by Claude itself. In the next section, we’ll share some of what we learned from this process. Principles for writing effective tools In this section, we distill our learnings into a few guiding principles for writing effective tools. Choosing the right tools for agents More tools don’t always lead to better outcomes. A common error we’ve observed is tools that merely wrap existing software functionality or API endpoints—whether or not the tools are appropriate for agents. This is because agents have distinct “affordances” to traditional software—that is, they have different ways of perceiving the potential actions they can take with those tools LLM agents have limited "context" (that is, there are limits to how much information they can process at once), whereas computer memory is cheap and abundant. Consider the task of searching for a contact in an address book. Traditional software programs can efficiently store and process a list of contacts one at a time, checking each one before moving on. However, if an LLM agent uses a tool that returns ALL contacts and then has to read through each one token-by-token, it's wasting its limited context space on irrelevant information (imagine searching for a contact in your address book by reading each page from top-to-bottom—that is, via brute-force search). The better and more natural approach (for agents and humans alike) is to skip to the relevant page first (perhaps finding it alphabetically). We recommend building a few thoughtful tools targeting specific high-impact workflows, which match your evaluation tasks and scaling up from there. In the address book case, you might choose to implement a search_contacts or message_contact tool instead of a list_contacts tool. Tools can consolidate functionality, handling potentially multiple discrete operations (or API calls) under the hood. For example, tools can enrich tool responses with related metadata or handle frequently chained, multi-step tasks in a single tool call. Here are some examples: Instead of implementing a list_users, list_events, and create_event tools, consider implementing a schedule_event tool which finds availability and schedules an event. Instead of implementing a read_logs tool, consider implementing a search_logs tool which only returns relevant log lines and some surrounding context. Instead of implementing get_customer_by_id, list_transactions, and list_notes tools, implement a get_customer_context tool which compiles all of a customer’s recent & relevant information all at once. Make sure each tool you build has a clear, distinct purpose. Tools should enable agents to subdivide and solve tasks in much the same way that a human would, given access to the same underlying resources, and simultaneously reduce the context that would have otherwise been consumed by intermediate outputs. Too many tools or overlapping tools can also distract agents from pursuing efficient strategies. Careful, selective planning of the tools you build (or don’t build) can really pay off. Namespacing your tools Your AI agents will potentially gain access to dozens of MCP servers and hundreds of different tools–including those by other developers. When tools overlap in function or have a vague purpose, agents can get confused about which ones to use. Namespacing (grouping related tools under common prefixes) can help delineate boundaries between lots of tools; MCP clients sometimes do this by default. For example, namespacing tools by service (e.g., asana_search, jira_search) and by resource (e.g., asana_projects_search, asana_users_search), can help agents select the right tools at the right time. We have found selecting between prefix- and suffix-based namespacing to have non-trivial effects on our tool-use evaluations. Effects vary by LLM and we encourage you to choose a naming scheme according to your own evaluations. Agents might call the wrong tools, call the right tools with the wrong parameters, call too few tools, or process tool responses incorrectly. By selectively implementing tools whose names reflect natural subdivisions of tasks, you simultaneously reduce the number of tools and tool descriptions loaded into the agent’s context and offload agentic computation from the agent’s context back into the tool calls themselves. This reduces an agent’s overall risk of making mistakes. Returning meaningful context from your tools In the same vein, tool implementations should take care to return only high signal information back to agents. They should prioritize contextual relevance over flexibility, and eschew low-level technical identifiers (for example: uuid, 256px_image_url, mime_type). Fields like name, image_url, and file_type are much more likely to directly inform agents’ downstream actions and responses. Agents also tend to grapple with natural language names, terms, or identifiers significantly more successfully than they do with cryptic identifiers. We’ve found that merely resolving arbitrary alphanumeric UUIDs to more semantically meaningful and interpretable language (or even a 0-indexed ID scheme) significantly improves Claude’s precision in retrieval tasks by reducing hallucinations. In some instances, agents may require the flexibility to interact with both natural language and technical identifiers outputs, if only to trigger downstream tool calls (for example, search_user(name=’jane’) → send_message(id=12345)). You can enable both by exposing a simple response_format enum parameter in your tool, allowing your agent to control whether tools return “concise” or “detailed” responses (images below). You can add more formats for even greater flexibility, similar to GraphQL where you can choose exactly which pieces of information you want to receive. Here is an example ResponseFormat enum to control tool response verbosity: enum ResponseFormat { DETAILED = "detailed", CONCISE = "concise" } Copy Here’s an example of a detailed tool response (206 tokens): This code snippet depicts an example of a detailed tool response. Here’s an example of a concise tool response (72 tokens): This code snippet depicts a concise tool response. Slack threads and thread replies are identified by unique thread_ts which are required to fetch thread replies. thread_ts and other IDs (channel_id, user_id) can be retrieved from a “detailed” tool response to enable further tool calls that require these. “concise” tool responses return only thread content and exclude IDs. In this example, we use ~⅓ of the tokens with “concise” tool responses. Even your tool response structure—for example XML, JSON, or Markdown—can have an impact on evaluation performance: there is no one-size-fits-all solution. This is because LLMs are trained on next-token prediction and tend to perform better with formats that match their training data. The optimal response structure will vary widely by task and agent. We encourage you to select the best response structure based on your own evaluation. Optimizing tool responses for token efficiency Optimizing the quality of context is important. But so is optimizing the quantity of context returned back to agents in tool responses. We suggest implementing some combination of pagination, range selection, filtering, and/or truncation with sensible default parameter values for any tool responses that could use up lots of context. For Claude Code, we restrict tool responses to 25,000 tokens by default. We expect the effective context length of agents to grow over time, but the need for context-efficient tools to remain. If you choose to truncate responses, be sure to steer agents with helpful instructions. You can directly encourage agents to pursue more token-efficient strategies, like making many small and targeted searches instead of a single, broad search for a knowledge retrieval task. Similarly, if a tool call raises an error (for example, during input validation), you can prompt-engineer your error responses to clearly communicate specific and actionable improvements, rather than opaque error codes or tracebacks. Here’s an example of a truncated tool response: This image depicts an example of a truncated tool response. Here’s an example of an unhelpful error response: This image depicts an example of an unhelpful tool response. Here’s an example of a helpful error response: This image depicts an example of a helpful error response. Tool truncation and error responses can steer agents towards more token-efficient tool-use behaviors (using filters or pagination) or give examples of correctly formatted tool inputs. Prompt-engineering your tool descriptions We now come to one of the most effective methods for improving tools: prompt-engineering your tool descriptions and specs. Because these are loaded into your agents’ context, they can collectively steer agents toward effective tool-calling behaviors. When writing tool descriptions and specs, think of how you would describe your tool to a new hire on your team. Consider the context that you might implicitly bring—specialized query formats, definitions of niche terminology, relationships between underlying resources—and make it explicit. Avoid ambiguity by clearly describing (and enforcing with strict data models) expected inputs and outputs. In particular, input parameters should be unambiguously named: instead of a parameter named user, try a parameter named user_id. With your evaluation you can measure the impact of your prompt engineering with greater confidence. Even small refinements to tool descriptions can yield dramatic improvements. Claude Sonnet 3.5 achieved state-of-the-art performance on the SWE-bench Verified evaluation after we made precise refinements to tool descriptions, dramatically reducing error rates and improving task completion. You can find other best practices for tool definitions in our Developer Guide. If you’re building tools for Claude, we also recommend reading about how tools are dynamically loaded into Claude’s system prompt. Lastly, if you’re writing tools for an MCP server, tool annotations help disclose which tools require open-world access or make destructive changes. Looking ahead To build effective tools for agents, we need to re-orient our software development practices from predictable, deterministic patterns to non-deterministic ones. Through the iterative, evaluation-driven process we’ve described in this post, we've identified consistent patterns in what makes tools successful: Effective tools are intentionally and clearly defined, use agent context judiciously, can be combined together in diverse workflows, and enable agents to intuitively solve real-world tasks. In the future, we expect the specific mechanisms through which agents interact with the world to evolve—from updates to the MCP protocol to upgrades to the underlying LLMs themselves. With a systematic, evaluation-driven approach to improving tools for agents, we can ensure that as agents become more capable, the tools they use will evolve alongside them. Acknowledgements Written by Ken Aizawa with valuable contributions from colleagues across Research (Barry Zhang, Zachary Witten, Daniel Jiang, Sami Al-Sheikh, Matt Bell, Maggie Vo), MCP (Theodora Chu, John Welsh, David Soria Parra, Adam Jones), Product Engineering (Santiago Seira), Marketing (Molly Vorwerck), Design (Drew Roper), and Applied AI (Christian Ryan, Alexander Bricken). 1Beyond training the underlying LLMs themselves. ------- Effective context engineering for AI agents Published Sep 29, 2025 Context is a critical but finite resource for AI agents. In this post, we explore strategies for effectively curating and managing the context that powers them. After a few years of prompt engineering being the focus of attention in applied AI, a new term has come to prominence: context engineering. Building with language models is becoming less about finding the right words and phrases for your prompts, and more about answering the broader question of “what configuration of context is most likely to generate our model’s desired behavior?" Context refers to the set of tokens included when sampling from a large-language model (LLM). The engineering problem at hand is optimizing the utility of those tokens against the inherent constraints of LLMs in order to consistently achieve a desired outcome. Effectively wrangling LLMs often requires thinking in context — in other words: considering the holistic state available to the LLM at any given time and what potential behaviors that state might yield. In this post, we’ll explore the emerging art of context engineering and offer a refined mental model for building steerable, effective agents. Context engineering vs. prompt engineering At Anthropic, we view context engineering as the natural progression of prompt engineering. Prompt engineering refers to methods for writing and organizing LLM instructions for optimal outcomes (see our docs for an overview and useful prompt engineering strategies). Context engineering refers to the set of strategies for curating and maintaining the optimal set of tokens (information) during LLM inference, including all the other information that may land there outside of the prompts. In the early days of engineering with LLMs, prompting was the biggest component of AI engineering work, as the majority of use cases outside of everyday chat interactions required prompts optimized for one-shot classification or text generation tasks. As the term implies, the primary focus of prompt engineering is how to write effective prompts, particularly system prompts. However, as we move towards engineering more capable agents that operate over multiple turns of inference and longer time horizons, we need strategies for managing the entire context state (system instructions, tools, Model Context Protocol (MCP), external data, message history, etc). An agent running in a loop generates more and more data that could be relevant for the next turn of inference, and this information must be cyclically refined. Context engineering is the art and science of curating what will go into the limited context window from that constantly evolving universe of possible information. Prompt engineering vs. context engineering In contrast to the discrete task of writing a prompt, context engineering is iterative and the curation phase happens each time we decide what to pass to the model. Why context engineering is important to building capable agents Despite their speed and ability to manage larger and larger volumes of data, we’ve observed that LLMs, like humans, lose focus or experience confusion at a certain point. Studies on needle-in-a-haystack style benchmarking have uncovered the concept of context rot: as the number of tokens in the context window increases, the model’s ability to accurately recall information from that context decreases. While some models exhibit more gentle degradation than others, this characteristic emerges across all models. Context, therefore, must be treated as a finite resource with diminishing marginal returns. Like humans, who have limited working memory capacity, LLMs have an “attention budget” that they draw on when parsing large volumes of context. Every new token introduced depletes this budget by some amount, increasing the need to carefully curate the tokens available to the LLM. This attention scarcity stems from architectural constraints of LLMs. LLMs are based on the transformer architecture, which enables every token to attend to every other token across the entire context. This results in n² pairwise relationships for n tokens. As its context length increases, a model's ability to capture these pairwise relationships gets stretched thin, creating a natural tension between context size and attention focus. Additionally, models develop their attention patterns from training data distributions where shorter sequences are typically more common than longer ones. This means models have less experience with, and fewer specialized parameters for, context-wide dependencies. Techniques like position encoding interpolation allow models to handle longer sequences by adapting them to the originally trained smaller context, though with some degradation in token position understanding. These factors create a performance gradient rather than a hard cliff: models remain highly capable at longer contexts but may show reduced precision for information retrieval and long-range reasoning compared to their performance on shorter contexts. These realities mean that thoughtful context engineering is essential for building capable agents. The anatomy of effective context Given that LLMs are constrained by a finite attention budget, good context engineering means finding the smallest possible set of high-signal tokens that maximize the likelihood of some desired outcome. Implementing this practice is much easier said than done, but in the following section, we outline what this guiding principle means in practice across the different components of context. System prompts should be extremely clear and use simple, direct language that presents ideas at the right altitude for the agent. The right altitude is the Goldilocks zone between two common failure modes. At one extreme, we see engineers hardcoding complex, brittle logic in their prompts to elicit exact agentic behavior. This approach creates fragility and increases maintenance complexity over time. At the other extreme, engineers sometimes provide vague, high-level guidance that fails to give the LLM concrete signals for desired outputs or falsely assumes shared context. The optimal altitude strikes a balance: specific enough to guide behavior effectively, yet flexible enough to provide the model with strong heuristics to guide behavior. Calibrating the system prompt in the process of context engineering. At one end of the spectrum, we see brittle if-else hardcoded prompts, and at the other end we see prompts that are overly general or falsely assume shared context. We recommend organizing prompts into distinct sections (like , , ## Tool guidance, ## Output description, etc) and using techniques like XML tagging or Markdown headers to delineate these sections, although the exact formatting of prompts is likely becoming less important as models become more capable. Regardless of how you decide to structure your system prompt, you should be striving for the minimal set of information that fully outlines your expected behavior. (Note that minimal does not necessarily mean short; you still need to give the agent sufficient information up front to ensure it adheres to the desired behavior.) It’s best to start by testing a minimal prompt with the best model available to see how it performs on your task, and then add clear instructions and examples to improve performance based on failure modes found during initial testing. Tools allow agents to operate with their environment and pull in new, additional context as they work. Because tools define the contract between agents and their information/action space, it’s extremely important that tools promote efficiency, both by returning information that is token efficient and by encouraging efficient agent behaviors. In Writing tools for AI agents – with AI agents, we discussed building tools that are well understood by LLMs and have minimal overlap in functionality. Similar to the functions of a well-designed codebase, tools should be self-contained, robust to error, and extremely clear with respect to their intended use. Input parameters should similarly be descriptive, unambiguous, and play to the inherent strengths of the model. One of the most common failure modes we see is bloated tool sets that cover too much functionality or lead to ambiguous decision points about which tool to use. If a human engineer can’t definitively say which tool should be used in a given situation, an AI agent can’t be expected to do better. As we’ll discuss later, curating a minimal viable set of tools for the agent can also lead to more reliable maintenance and pruning of context over long interactions. Providing examples, otherwise known as few-shot prompting, is a well known best practice that we continue to strongly advise. However, teams will often stuff a laundry list of edge cases into a prompt in an attempt to articulate every possible rule the LLM should follow for a particular task. We do not recommend this. Instead, we recommend working to curate a set of diverse, canonical examples that effectively portray the expected behavior of the agent. For an LLM, examples are the “pictures” worth a thousand words. Our overall guidance across the different components of context (system prompts, tools, examples, message history, etc) is to be thoughtful and keep your context informative, yet tight. Now let's dive into dynamically retrieving context at runtime. Context retrieval and agentic search In Building effective AI agents, we highlighted the differences between LLM-based workflows and agents. Since we wrote that post, we’ve gravitated towards a simple definition for agents: LLMs autonomously using tools in a loop. Working alongside our customers, we’ve seen the field converging on this simple paradigm. As the underlying models become more capable, the level of autonomy of agents can scale: smarter models allow agents to independently navigate nuanced problem spaces and recover from errors. We’re now seeing a shift in how engineers think about designing context for agents. Today, many AI-native applications employ some form of embedding-based pre-inference time retrieval to surface important context for the agent to reason over. As the field transitions to more agentic approaches, we increasingly see teams augmenting these retrieval systems with “just in time” context strategies. Rather than pre-processing all relevant data up front, agents built with the “just in time” approach maintain lightweight identifiers (file paths, stored queries, web links, etc.) and use these references to dynamically load data into context at runtime using tools. Anthropic’s agentic coding solution Claude Code uses this approach to perform complex data analysis over large databases. The model can write targeted queries, store results, and leverage Bash commands like head and tail to analyze large volumes of data without ever loading the full data objects into context. This approach mirrors human cognition: we generally don’t memorize entire corpuses of information, but rather introduce external organization and indexing systems like file systems, inboxes, and bookmarks to retrieve relevant information on demand. Beyond storage efficiency, the metadata of these references provides a mechanism to efficiently refine behavior, whether explicitly provided or intuitive. To an agent operating in a file system, the presence of a file named test_utils.py in a tests folder implies a different purpose than a file with the same name located in src/core_logic/ Folder hierarchies, naming conventions, and timestamps all provide important signals that help both humans and agents understand how and when to utilize information. Letting agents navigate and retrieve data autonomously also enables progressive disclosure—in other words, allows agents to incrementally discover relevant context through exploration. Each interaction yields context that informs the next decision: file sizes suggest complexity; naming conventions hint at purpose; timestamps can be a proxy for relevance. Agents can assemble understanding layer by layer, maintaining only what's necessary in working memory and leveraging note-taking strategies for additional persistence. This self-managed context window keeps the agent focused on relevant subsets rather than drowning in exhaustive but potentially irrelevant information. Of course, there's a trade-off: runtime exploration is slower than retrieving pre-computed data. Not only that, but opinionated and thoughtful engineering is required to ensure that an LLM has the right tools and heuristics for effectively navigating its information landscape. Without proper guidance, an agent can waste context by misusing tools, chasing dead-ends, or failing to identify key information. In certain settings, the most effective agents might employ a hybrid strategy, retrieving some data up front for speed, and pursuing further autonomous exploration at its discretion. The decision boundary for the ‘right’ level of autonomy depends on the task. Claude Code is an agent that employs this hybrid model: CLAUDE.md files are naively dropped into context up front, while primitives like glob and grep allow it to navigate its environment and retrieve files just-in-time, effectively bypassing the issues of stale indexing and complex syntax trees. The hybrid strategy might be better suited for contexts with less dynamic content, such as legal or finance work. As model capabilities improve, agentic design will trend towards letting intelligent models act intelligently, with progressively less human curation. Given the rapid pace of progress in the field, "do the simplest thing that works" will likely remain our best advice for teams building agents on top of Claude. Context engineering for long-horizon tasks Long-horizon tasks require agents to maintain coherence, context, and goal-directed behavior over sequences of actions where the token count exceeds the LLM’s context window. For tasks that span tens of minutes to multiple hours of continuous work, like large codebase migrations or comprehensive research projects, agents require specialized techniques to work around the context window size limitation. Waiting for larger context windows might seem like an obvious tactic. But it's likely that for the foreseeable future, context windows of all sizes will be subject to context pollution and information relevance concerns—at least for situations where the strongest agent performance is desired. To enable agents to work effectively across extended time horizons, we've developed a few techniques that address these context pollution constraints directly: compaction, structured note-taking, and multi-agent architectures. Compaction Compaction is the practice of taking a conversation nearing the context window limit, summarizing its contents, and reinitiating a new context window with the summary. Compaction typically serves as the first lever in context engineering to drive better long-term coherence. At its core, compaction distills the contents of a context window in a high-fidelity manner, enabling the agent to continue with minimal performance degradation. In Claude Code, for example, we implement this by passing the message history to the model to summarize and compress the most critical details. The model preserves architectural decisions, unresolved bugs, and implementation details while discarding redundant tool outputs or messages. The agent can then continue with this compressed context plus the five most recently accessed files. Users get continuity without worrying about context window limitations. The art of compaction lies in the selection of what to keep versus what to discard, as overly aggressive compaction can result in the loss of subtle but critical context whose importance only becomes apparent later. For engineers implementing compaction systems, we recommend carefully tuning your prompt on complex agent traces. Start by maximizing recall to ensure your compaction prompt captures every relevant piece of information from the trace, then iterate to improve precision by eliminating superfluous content. An example of low-hanging superfluous content is clearing tool calls and results – once a tool has been called deep in the message history, why would the agent need to see the raw result again? One of the safest lightest touch forms of compaction is tool result clearing, most recently launched as a feature on the Claude Developer Platform. Structured note-taking Structured note-taking, or agentic memory, is a technique where the agent regularly writes notes persisted to memory outside of the context window. These notes get pulled back into the context window at later times. This strategy provides persistent memory with minimal overhead. Like Claude Code creating a to-do list, or your custom agent maintaining a NOTES.md file, this simple pattern allows the agent to track progress across complex tasks, maintaining critical context and dependencies that would otherwise be lost across dozens of tool calls. Claude playing Pokémon demonstrates how memory transforms agent capabilities in non-coding domains. The agent maintains precise tallies across thousands of game steps—tracking objectives like "for the last 1,234 steps I've been training my Pokémon in Route 1, Pikachu has gained 8 levels toward the target of 10." Without any prompting about memory structure, it develops maps of explored regions, remembers which key achievements it has unlocked, and maintains strategic notes of combat strategies that help it learn which attacks work best against different opponents. After context resets, the agent reads its own notes and continues multi-hour training sequences or dungeon explorations. This coherence across summarization steps enables long-horizon strategies that would be impossible when keeping all the information in the LLM’s context window alone. As part of our Sonnet 4.5 launch, we released a memory tool in public beta on the Claude Developer Platform that makes it easier to store and consult information outside the context window through a file-based system. This allows agents to build up knowledge bases over time, maintain project state across sessions, and reference previous work without keeping everything in context. Sub-agent architectures Sub-agent architectures provide another way around context limitations. Rather than one agent attempting to maintain state across an entire project, specialized sub-agents can handle focused tasks with clean context windows. The main agent coordinates with a high-level plan while subagents perform deep technical work or use tools to find relevant information. Each subagent might explore extensively, using tens of thousands of tokens or more, but returns only a condensed, distilled summary of its work (often 1,000-2,000 tokens). This approach achieves a clear separation of concerns—the detailed search context remains isolated within sub-agents, while the lead agent focuses on synthesizing and analyzing the results. This pattern, discussed in How we built our multi-agent research system, showed a substantial improvement over single-agent systems on complex research tasks. The choice between these approaches depends on task characteristics. For example: Compaction maintains conversational flow for tasks requiring extensive back-and-forth; Note-taking excels for iterative development with clear milestones; Multi-agent architectures handle complex research and analysis where parallel exploration pays dividends. Even as models continue to improve, the challenge of maintaining coherence across extended interactions will remain central to building more effective agents. Conclusion Context engineering represents a fundamental shift in how we build with LLMs. As models become more capable, the challenge isn't just crafting the perfect prompt—it's thoughtfully curating what information enters the model's limited attention budget at each step. Whether you're implementing compaction for long-horizon tasks, designing token-efficient tools, or enabling agents to explore their environment just-in-time, the guiding principle remains the same: find the smallest set of high-signal tokens that maximize the likelihood of your desired outcome. The techniques we've outlined will continue evolving as models improve. We're already seeing that smarter models require less prescriptive engineering, allowing agents to operate with more autonomy. But even as capabilities scale, treating context as a precious, finite resource will remain central to building reliable, effective agents. Get started with context engineering in the Claude Developer Platform today, and access helpful tips and best practices via our memory and context management cookbook. --------- ffective harnesses for long-running agents Published Nov 26, 2025 Agents still face challenges working across many context windows. We looked to human engineers for inspiration in creating a more effective harness for long-running agents. As AI agents become more capable, developers are increasingly asking them to take on complex tasks requiring work that spans hours, or even days. However, getting agents to make consistent progress across multiple context windows remains an open problem. The core challenge of long-running agents is that they must work in discrete sessions, and each new session begins with no memory of what came before. Imagine a software project staffed by engineers working in shifts, where each new engineer arrives with no memory of what happened on the previous shift. Because context windows are limited, and because most complex projects cannot be completed within a single window, agents need a way to bridge the gap between coding sessions. We developed a two-fold solution to enable the Claude Agent SDK to work effectively across many context windows: an initializer agent that sets up the environment on the first run, and a coding agent that is tasked with making incremental progress in every session, while leaving clear artifacts for the next session. You can find code examples in the accompanying quickstart. The long-running agent problem The Claude Agent SDK is a powerful, general-purpose agent harness adept at coding, as well as other tasks that require the model to use tools to gather context, plan, and execute. It has context management capabilities such as compaction, which enables an agent to work on a task without exhausting the context window. Theoretically, given this setup, it should be possible for an agent to continue to do useful work for an arbitrarily long time. However, compaction isn’t sufficient. Out of the box, even a frontier coding model like Opus 4.5 running on the Claude Agent SDK in a loop across multiple context windows will fall short of building a production-quality web app if it’s only given a high-level prompt, such as “build a clone of claude.ai.” Claude’s failures manifested in two patterns. First, the agent tended to try to do too much at once—essentially to attempt to one-shot the app. Often, this led to the model running out of context in the middle of its implementation, leaving the next session to start with a feature half-implemented and undocumented. The agent would then have to guess at what had happened, and spend substantial time trying to get the basic app working again. This happens even with compaction, which doesn’t always pass perfectly clear instructions to the next agent. A second failure mode would often occur later in a project. After some features had already been built, a later agent instance would look around, see that progress had been made, and declare the job done. This decomposes the problem into two parts. First, we need to set up an initial environment that lays the foundation for all the features that a given prompt requires, which sets up the agent to work step-by-step and feature-by-feature. Second, we should prompt each agent to make incremental progress towards its goal while also leaving the environment in a clean state at the end of a session. By “clean state” we mean the kind of code that would be appropriate for merging to a main branch: there are no major bugs, the code is orderly and well-documented, and in general, a developer could easily begin work on a new feature without first having to clean up an unrelated mess. When experimenting internally, we addressed these problems using a two-part solution: Initializer agent: The very first agent session uses a specialized prompt that asks the model to set up the initial environment: an init.sh script, a claude-progress.txt file that keeps a log of what agents have done, and an initial git commit that shows what files were added. Coding agent: Every subsequent session asks the model to make incremental progress, then leave structured updates.1 The key insight here was finding a way for agents to quickly understand the state of work when starting with a fresh context window, which is accomplished with the claude-progress.txt file alongside the git history. Inspiration for these practices came from knowing what effective software engineers do every day. Environment management In the updated Claude 4 prompting guide, we shared some best practices for multi-context window workflows, including a harness structure that uses “a different prompt for the very first context window.” This “different prompt” requests that the initializer agent set up the environment with all the necessary context that future coding agents will need to work effectively. Here, we provide a deeper dive on some of the key components of such an environment. Feature list To address the problem of the agent one-shotting an app or prematurely considering the project complete, we prompted the initializer agent to write a comprehensive file of feature requirements expanding on the user’s initial prompt. In the claude.ai clone example, this meant over 200 features, such as “a user can open a new chat, type in a query, press enter, and see an AI response.” These features were all initially marked as “failing” so that later coding agents would have a clear outline of what full functionality looked like. { "category": "functional", "description": "New chat button creates a fresh conversation", "steps": [ "Navigate to main interface", "Click the 'New Chat' button", "Verify a new conversation is created", "Check that chat area shows welcome state", "Verify conversation appears in sidebar" ], "passes": false } Copy We prompt coding agents to edit this file only by changing the status of a passes field, and we use strongly-worded instructions like “It is unacceptable to remove or edit tests because this could lead to missing or buggy functionality.” After some experimentation, we landed on using JSON for this, as the model is less likely to inappropriately change or overwrite JSON files compared to Markdown files. Incremental progress Given this initial environment scaffolding, the next iteration of the coding agent was then asked to work on only one feature at a time. This incremental approach turned out to be critical to addressing the agent’s tendency to do too much at once. Once working incrementally, it’s still essential that the model leaves the environment in a clean state after making a code change. In our experiments, we found that the best way to elicit this behavior was to ask the model to commit its progress to git with descriptive commit messages and to write summaries of its progress in a progress file. This allowed the model to use git to revert bad code changes and recover working states of the code base. These approaches also increased efficiency, as they eliminated the need for an agent to have to guess at what had happened and spend its time trying to get the basic app working again. Testing One final major failure mode that we observed was Claude’s tendency to mark a feature as complete without proper testing. Absent explicit prompting, Claude tended to make code changes, and even do testing with unit tests or curl commands against a development server, but would fail recognize that the feature didn’t work end-to-end. In the case of building a web app, Claude mostly did well at verifying features end-to-end once explicitly prompted to use browser automation tools and do all testing as a human user would. Screenshots taken by Claude through the Puppeteer MCP server as it tested the claude.ai clone. Screenshots taken by Claude through the Puppeteer MCP server as it tested the claude.ai clone. Providing Claude with these kinds of testing tools dramatically improved performance, as the agent was able to identify and fix bugs that weren’t obvious from the code alone. Some issues remain, like limitations to Claude’s vision and to browser automation tools making it difficult to identify every kind of bug. For example, Claude can’t see browser-native alert modals through the Puppeteer MCP, and features relying on these modals tended to be buggier as a result. Getting up to speed With all of the above in place, every coding agent is prompted to run through a series of steps to get its bearings, some quite basic but still helpful: Run pwd to see the directory you’re working in. You’ll only be able to edit files in this directory. Read the git logs and progress files to get up to speed on what was recently worked on. Read the features list file and choose the highest-priority feature that’s not yet done to work on. This approach saves Claude some tokens in every session since it doesn’t have to figure out how to test the code. It also helps to ask the initializer agent to write an init.sh script that can run the development server, and then run through a basic end-to-end test before implementing a new feature. In the case of the claude.ai clone, this meant that the agent always started the local development server and used the Puppeteer MCP to start a new chat, send a message, and receive a response. This ensured that Claude could quickly identify if the app had been left in a broken state, and immediately fix any existing bugs. If the agent had instead started implementing a new feature, it would likely make the problem worse. Given all this, a typical session starts off with the following assistant messages: [Assistant] I'll start by getting my bearings and understanding the current state of the project. [Tool Use] [Tool Use] [Tool Use] [Assistant] Let me check the git log to see recent work. [Tool Use] [Assistant] Now let me check if there's an init.sh script to restart the servers. [Assistant] Excellent! Now let me navigate to the application and verify that some fundamental features are still working. [Assistant] Based on my verification testing, I can see that the fundamental functionality is working well. The core chat features, theme switching, conversation loading, and error handling are all functioning correctly. Now let me review the tests.json file more comprehensively to understand what needs to be implemented next. Copy Agent failure modes and solutions Problem Initializer Agent Behavior Coding Agent Behavior Claude declares victory on the entire project too early. Set up a feature list file: based on the input spec, set up a structured JSON file with a list of end-to-end feature descriptions. Read the feature list file at the beginning of a session. Choose a single feature to start working on. Claude leaves the environment in a state with bugs or undocumented progress. An initial git repo and progress notes file is written. Start the session by reading the progress notes file and git commit logs, and run a basic test on the development server to catch any undocumented bugs. End the session by writing a git commit and progress update. Claude marks features as done prematurely. Set up a feature list file. Self-verify all features. Only mark features as “passing” after careful testing. Claude has to spend time figuring out how to run the app. Write an init.sh script that can run the development server. Start the session by reading init.sh. Summarizing four common failure modes and solutions in long-running AI agents. Future work This research demonstrates one possible set of solutions in a long-running agent harness to enable the model to make incremental progress across many context windows. However, there remain open questions. Most notably, it’s still unclear whether a single, general-purpose coding agent performs best across contexts, or if better performance can be achieved through a multi-agent architecture. It seems reasonable that specialized agents like a testing agent, a quality assurance agent, or a code cleanup agent, could do an even better job at sub-tasks across the software development lifecycle. Additionally, this demo is optimized for full-stack web app development. A future direction is to generalize these findings to other fields. It’s likely that some or all of these lessons can be applied to the types of long-running agentic tasks required in, for example, scientific research or financial modeling. ) --- ## File: docs/claude_research.md --- name: production-grade-llm-agents doc_type: research source_url: No --- Engineering Production-Grade LLM Agents: A Technical Deep Dive The shift from prompt engineering to context engineering represents the most significant paradigm change in building LLM agents. As Anthropic's research articulates, the challenge isn't writing better prompts—it's curating "the smallest possible set of high-signal tokens that maximize the likelihood of desired outcomes." Inkeepanthropic This report synthesizes technical findings from major AI labs and framework developers on multi-agent architectures, context management, attention degradation, and agent reliability patterns. Multi-agent architectures: From orchestrators to swarms Production multi-agent systems have converged on three dominant patterns, each with distinct tradeoffs. Orchestrator-worker (supervisor) patterns place a central agent in control, delegating to specialists and synthesizing results. LangGraph's benchmarks found this architecture initially performed 50% worse than optimized versions due to the "telephone game" problem—supervisors paraphrasing sub-agent responses incorrectly. The fix: implementing a forward_message tool allowing sub-agents to pass responses directly to users. langchainLangChain Swarm architectures, pioneered by OpenAI's experimental Swarm framework, enable peer-to-peer handoffs where any agent transfers control to any other. LangGraph benchmarks show swarms slightly outperform supervisors because sub-agents respond directly to users, eliminating translation errors. langchainLangChain The core abstraction is elegantly simple: pythondef transfer_to_agent_b(): return agent_b # Handoff via function return agent_a = Agent( name="Agent A", functions=[transfer_to_agent_b] ) Hierarchical patterns, implemented in CrewAI's Process.hierarchical mode, create management trees where managers decompose goals and delegate to subordinates. Activewizards This mirrors organizational structures and works well for complex, multi-stage tasks. The critical insight from Manus AI's production experience: sub-agents exist primarily to isolate context, not to anthropomorphize role division. Rlancemartin Context isolation prevents KV-cache penalties and avoids context confusion between specialized tasks. Context coordination and the file system as memory How agents share context determines both performance and cost. Manus AI identified KV-cache hit rate as the single most important production metric— Manusthe difference between $0.30/MTok (cached) and $3/MTok (uncached) for Claude Sonnet, a 10× cost differential. manus Three context-sharing patterns emerge from production systems: PatternMechanismUse CaseFull context delegationPlanner shares entire context with sub-agentComplex tasks requiring complete understandingInstruction passingPlanner creates instructions via function callSimple, well-defined subtasksFile system memoryAgents read/write to persistent storageUnlimited size, agent-operable context Claude Code exemplifies file-system-as-memory: rather than stuffing context windows, agents use grep, head, and tail to navigate codebases, storing query results and analyzing large databases without loading full data. AnthropicRlancemartin This "just-in-time" context loading maintains small active context while enabling access to arbitrarily large information. anthropic Manus AI's context engineering principles offer production-tested guidance: use append-only context (never modify previous actions), employ logit masking instead of tool removal to constrain actions, and keep errors in context for implicit belief updates rather than hiding failures. manusManus KV-cache optimization: From PagedAttention to prefix caching The KV-cache stores Key and Value tensors computed during inference, growing linearly with sequence length. Neptune.ai For LLaMA-2 13B, this means approximately 1MB per token per sequence—a 4K context consumes ~4GB, comparable to the model itself. Rohan-paul PagedAttention, introduced by vLLM, revolutionized memory efficiency by applying OS-inspired virtual memory concepts. Medium Instead of pre-allocating contiguous memory, it partitions KV cache into fixed-size blocks (typically 16 tokens), mapping logical blocks to non-contiguous physical memory via block tables. Results: 2-4× throughput improvement arXiv with up to 96% reduction in memory waste. Medium Prefix caching (Automatic Prefix Caching) reuses KV blocks across requests sharing identical prefixes, using hash-based block matching: hash(parent_hash, block_tokens, extra_hashes). Anthropic reports up to 90% cost savings and 85% latency reduction with prefix caching on Claude. Advanced quantization pushes efficiency further. SKVQ achieves 1M token context on 80GB GPUs using 2-bit keys and 1.5-bit values with only <5% accuracy drop. Emergent Mind Layer-Condensed KV caches only top layers for 26× throughput. Emergent Mind RazorAttention identifies "retrieval heads" that need full caches versus those that can use buffers, achieving 40-60% memory reduction. Emergent Mind Context rot: The hidden performance cliff Despite claims of 100K+ token context windows, empirical research reveals significant performance degradation—a phenomenon researchers call context rot. anthropic The "lost in the middle" effect, documented by Liu et al. (TACL 2024), shows a U-shaped performance curve: accuracy drops 10-40% when relevant information sits in the middle of context versus beginning or end. arXivACL Anthology The RULER benchmark delivers a sobering finding: only half of models claiming 32K+ context maintain satisfactory performance at 32K tokens. arXivOpenReview GPT-4 showed the least degradation (15.4 points from 4K to 128K), while most models dropped 30+ points. Medium Near-perfect scores on simple needle-in-haystack tests don't translate to real long-context understanding— trychromaRULER's multi-hop tracing, aggregation, and question-answering tasks expose the gap. arXivOpenReview Chroma's 2025 research across 18 LLMs identified critical patterns: trychroma Distractor effect: Even a single irrelevant document reduces performance; multiple distractors compound degradation Needle-question similarity: Lower similarity pairs show faster degradation with context length trychroma Counterintuitive haystack structure: Shuffled (incoherent) haystacks produce better performance than logically coherent ones trychroma Model-specific behaviors: Claude shows lowest hallucination rates but high abstention under ambiguity; GPT shows highest hallucination rates with confident-but-incorrect responses trychroma Four failure modes in production contexts Beyond simple degradation, long-running agents encounter distinct context failure patterns that require different mitigations: Context poisoning occurs when hallucinations or errors enter context and compound through repeated reference. Feluda As Drew Breunig documents, if an agent's "goals" section becomes poisoned, it develops nonsensical strategies that take "very long time to undo." Drew Breunig Symptoms include degraded output quality, tool misalignment, and hallucinations treated as facts. Context distraction emerges when context grows so long that models over-focus on context at the expense of training knowledge. The Gemini 2.5 technical report notes: "While Gemini 2.5 Pro supports 1M+ token context, making effective use of it for agents presents a new research frontier." Drew Breunig Context confusion arises when irrelevant information influences responses. As one practitioner observed: "If you put something in the context, the model has to pay attention to it. It may be irrelevant information or needless tool definitions, but the model will take it into account." Drew Breunig Context clash develops when accumulated information directly conflicts, documented by Microsoft and Salesforce research showing that sharding information across multiple prompts creates conflicting contexts that derail reasoning. Drew Breunig Mitigation strategies that work Effective context management employs four strategies, formalized by LangChain as the "four-bucket" approach: StrategyImplementationExampleWriteSave context outside windowScratchpads, memory stores, file systemSelectPull relevant context inRAG, memory retrieval, tool selectionCompressReduce tokens preserving infoSummarization, observation maskingIsolateSplit context across agentsSub-agents, sandboxes, state schemas Observation masking deserves special attention: replacing old tool outputs with fixed masks like "Previous X lines elided for brevity" often matches or exceeds LLM summarization performance while adding zero token overhead (versus 5-7% for summarization). Research shows observations comprise 83.9% of tokens in typical agent trajectories—masking offers significant efficiency gains. Architectural approaches include Core Context Aware (CCA) Attention, a plug-and-play module achieving 5.7× faster inference at 64K tokens, arXiv and Google's Chain of Agents (CoA), which breaks inputs into chunks processed by worker agents sequentially, reducing time complexity from n² to nk. Google Research Tool design for agent ergonomics Tools are contracts between deterministic systems and non-deterministic agents—design matters critically. anthropic Anthropic's guidance emphasizes minimizing functional overlap: "If a human can't definitively say which tool to use, an AI agent can't either." Anthropic The consolidation principle transforms API design: Instead ofImplementlist_users, list_events, create_eventschedule_event (finds availability + schedules)read_logssearch_logs (returns relevant lines with context)get_customer_by_id, list_transactions, list_notesget_customer_context (compiles all relevant info) Tool descriptions require engineering. Poor descriptions like "Search the database" with cryptic parameter names force agents to guess. Optimized descriptions include usage context ("Use this when the user asks about company policies"), examples ("Example: 'vacation policy remote employees'"), and defaults ("Start with 3-5 for most queries"). Response format options offer significant token savings: implementing a response_format parameter with DETAILED (full JSON, 206 tokens) versus CONCISE (essential info only, 72 tokens) cuts context consumption by 65% when full metadata isn't needed. Reasoning patterns and their measured impact ReAct (Reasoning + Acting) interleaves thinking with tool use: "Thought 1: [reasoning] → Action 1: [tool call] → Observation 1: [result]". Prompt Engineering Guide Performance gains are substantial: +34% absolute success rate on ALFWorld, +10% on WebShop versus imitation learning. React-lm However, 2024 research reveals brittleness—40-90% of generated thoughts lead to invalid actions depending on the model. arXiv Tree of Thoughts (ToT) explores multiple reasoning paths simultaneously. On Game of 24, performance jumps from 4% (Chain-of-Thought) to 74% with GPT-4 using ToT. KDnuggets The approach works by generating multiple candidates at each reasoning step, having the LLM self-evaluate progress, and using tree search (BFS/DFS) for exploration. Dynamic few-shot selection consistently outperforms static examples. LangChain benchmarks show Claude 3 Sonnet jumping from 16% to 52% accuracy with just 3 semantically similar examples—often matching or exceeding 13 static examples. The key is semantic similarity: retrieve examples similar to the current query rather than maintaining fixed lists. Hallucination prevention in agentic contexts Agentic settings amplify hallucination risk since errors compound across tool calls. A critical MIT survey finding: "No prior work demonstrates successful self-correction with feedback from prompted LLMs, except for tasks exceptionally suited for self-correction." What does work for self-correction: External tool feedback: Code execution results, API verification, calculator outputs Retrieval grounding: Web search for fact verification Fine-tuned correction models: Models specifically trained for correction tasks RAG-based grounding can decrease hallucination by 60-80% according to industry surveys. Implementation requires explicit constraints: "Answer based ONLY on the provided context. If the context doesn't contain relevant information, respond: 'I cannot find information about this in the provided documents.'" The Chain-of-Verification (CoVe) pattern generates verification questions about claims, answers them independently, compares answers with initial claims, and revises based on inconsistencies. ProCo framework achieves +6.8 EM on QA and +14.1% on arithmetic through systematic condition verification. Evaluation methods for production agents Anthropic's multi-agent evaluation approach uses a structured rubric: factual accuracy (claims match sources), citation accuracy (cited sources match claims), completeness (all aspects covered), source quality (primary versus secondary), and tool efficiency (reasonable usage). Anthropic Key benchmarks reveal capability gaps: BenchmarkFindingRULEROnly 50% of 32K+ models maintain performance at 32K tokens arXiv∞Bench"Existing long-context LLMs require significant advancements for 100K+"LongBench v2Best model achieves 50.1% accuracy; humans achieve 53.7% Longbench2τ-benchTests single/multi-agent cognitive architectures on real-world scenarios The methodology: start with small samples (~20 queries), use LLM-as-judge for scalable evaluation, supplement with human evaluation to catch automation misses, and focus on end-state evaluation for agents that mutate state. Anthropic Conclusion Building production LLM agents requires treating context as the central engineering concern rather than an afterthought. The research converges on several principles: Context quality trumps context length—despite 1M+ token windows, effective performance often degrades past 32K-256K tokens depending on task complexity. Use just-in-time context loading, observation masking, and sub-agent isolation to maintain signal quality. Multi-agent architecture selection depends on coordination needs: swarms for peer-to-peer handoffs with direct user interaction, supervisors for integrating diverse sub-agents with minimal assumptions, hierarchical patterns for complex decomposition tasks. Tool design directly impacts agent capability. Consolidate overlapping tools, return contextual information in error messages, implement response format options, and namespace clearly. anthropic Poor tool descriptions create failure modes no amount of prompt engineering can fix. Verification requires external grounding. Self-correction without external feedback doesn't work reliably. RAG, tool execution results, and multi-agent verification architectures provide the grounding necessary for production reliability. The field is rapidly evolving—KV-cache optimization, attention architectures, and evaluation methods continue advancing. Engineers building agents should monitor production metrics (especially KV-cache hit rates and token efficiency), implement compaction triggers at 80% of effective context limits, and design systems assuming context will degrade rather than hoping it won't. --- ## File: docs/compression.md --- name: context-compression-evaluation doc_type: research source_url: No --- Evaluating Context Compression for AI Agents By Factory Research - December 16, 2025 - 10 minute read - Share Engineering Research New We built an evaluation framework to measure how much context different compression strategies preserve. After testing three approaches on real-world, long-running agent sessions spanning debugging, code review, and feature implementation, we found that structured summarization retains more useful information than alternatives from OpenAI and Anthropic. Table of Contents 01 The problem 02 Measuring context quality 03 Three approaches to compression 04 A concrete example 05 How the LLM judge works 06 Results 07 What we learned 08 Methodology details 09 Appendix: LLM Judge Prompts and Rubrics Tasteful abstract illustration evocative of memory and blurriness When an AI agent helps you work through a complex task across hundreds of messages, what happens when it runs out of memory? The answer determines whether your agent continues productively or starts asking "wait, what were we trying to do again?" We built an evaluation framework to measure how much context different compression strategies preserve. After testing three approaches on real-world, long-running agent sessions (debugging, PR review, feature implementation, CI troubleshooting, data science, ML research), we found that structured summarization retains more useful information than alternative methods from OpenAI and Anthropic, without sacrificing compression efficiency. Bar chart comparing quality scores by dimension across Factory, OpenAI, and Anthropic This post walks through the problem, our methodology, concrete examples of how different approaches perform, and what the results mean for building reliable AI agents. The problem Long-running agent sessions can generate millions of tokens of conversation history. That far exceeds what any model can hold in working memory. The naive solution is aggressive compression: squeeze everything into the smallest possible summary. But this increases the chance your agent forgets which files it modified or what approach it already tried. It is likely to waste tokens re-reading files and re-exploring dead ends. The right optimization target is not tokens per request. It is tokens per task. Measuring context quality Traditional metrics like ROUGE or embedding similarity do not tell you whether an agent can continue working effectively after compression. A summary might score high on lexical overlap while missing the one file path the agent needs to continue. We designed a probe-based evaluation that directly measures functional quality. The idea is simple: after compression, ask the agent questions that require remembering specific details from the truncated history. If the compression preserved the right information, the agent answers correctly. If not, it guesses or hallucinates. We use four probe types: Probe type What it tests Example question Recall Factual retention "What was the original error message?" Artifact File tracking "Which files have we modified? Describe what changed in each." Continuation Task planning "What should we do next?" Decision Reasoning chain "We discussed options for the Redis issue. What did we decide?" Recall probes test whether specific facts survive compression. Artifact probes test whether the agent knows what files it touched. Continuation probes test whether the agent can pick up where it left off. Decision probes test whether the reasoning behind past choices is preserved. We grade responses using an LLM judge (GPT-5.2) across six dimensions: Dimension What it measures Accuracy Are technical details correct? File paths, function names, errors Context awareness Does the response reflect current conversation state? Artifact trail Does the agent know which files were read or modified? Completeness Does the response address all parts of the question? Continuity Can work continue without re-fetching information? Instruction following Does the response follow the probe format? Each dimension is scored 0-5 using detailed rubrics. The rubrics specify what constitutes a 0 ("Completely fails"), 3 ("Adequately meets with minor issues"), and 5 ("Excellently meets with no issues") for each criterion. Why these dimensions matter for software development These dimensions were chosen specifically because they capture what goes wrong when coding agents lose context: Artifact trail is critical because coding agents need to know which files they have touched. Without this, an agent might re-read files it already examined, make conflicting edits, or lose track of test results. A ChatGPT conversation can afford to forget earlier topics; a coding agent that forgets it modified auth.controller.ts will produce inconsistent work. Continuity directly impacts token efficiency. When an agent cannot continue from where it left off, it re-fetches files and re-explores approaches it already tried. This wastes tokens and time, turning a single-pass task into an expensive multi-pass one. Context awareness matters because coding sessions have state. The agent needs to know not just facts from the past, but the current state of the task: what has been tried, what failed, what is left to do. Generic summarization often captures "what happened" while losing "where we are." Accuracy is non-negotiable for code. A wrong file path or misremembered function name leads to failed edits or hallucinated solutions. Unlike conversational AI where approximate recall is acceptable, coding agents need precise technical details. Completeness ensures the agent addresses all parts of a multi-part request. When a user asks to "fix the bug and add tests," a complete response handles both. Incomplete responses force follow-up prompts and waste tokens on re-establishing context. Instruction following verifies the agent respects constraints and formats. When asked to "only modify the auth module" or "output as JSON," the agent must comply. This dimension catches cases where compression preserved facts but lost the user's requirements. Three approaches to compression We compared three production-ready compression strategies. Factory maintains a structured, persistent summary with explicit sections for different information types: session intent, file modifications, decisions made, and next steps. When compression triggers, only the newly-truncated span is summarized and merged with the existing summary. We call this anchored iterative summarization. The key insight is that structure forces preservation. By dedicating sections to specific information types, the summary cannot silently drop file paths or skip over decisions. Each section acts as a checklist: the summarizer must populate it or explicitly leave it empty. This prevents the gradual information loss that occurs with freeform summarization. OpenAI offers the /responses/compact endpoint, which produces opaque, compressed representations optimized for reconstruction fidelity. It achieves the highest compression ratios (99.3%) but sacrifices interpretability. You cannot read the compressed output to verify what was preserved. Anthropic provides built-in context compression through the Claude SDK. It produces detailed, structured summaries (typically 7-12k characters) with sections for analysis, files, pending tasks, and current state. The key difference from Factory is the update mechanism: Anthropic regenerates the full summary on each compression, while Factory's anchored approach incrementally merges new information into a persistent summary. This affects both consistency across compressions and which details survive repeated compression cycles. A concrete example Consider a debugging session where a user reported a 401 error on the /api/auth/login endpoint despite valid credentials. Over 178 messages and 89,000 tokens, the assistant: Read the authentication controller and identified JWT token generation Examined the middleware chain and found a misconfigured CORS policy Discovered the session store was using an expired Redis connection Fixed the Redis connection pooling configuration Added retry logic for transient connection failures Ran the test suite (14 tests passing, 2 failing) Fixed the remaining test failures related to mock setup Files examined: auth.controller.ts, middleware/cors.ts, config/redis.ts, services/session.service.ts, tests/auth.test.ts After compression, we ask: "Before we continue, can you remind me what the original error was that started this debugging session? Please be specific about the error code and endpoint." Here is how each method responded: Factory response (scored 4.8/5): "The original error was a 401 Unauthorized response from the /api/auth/login endpoint. Users were receiving this error even when providing valid username/password combinations. The root cause turned out to be a stale Redis connection in the session store - the JWT was being generated correctly, but the session could not be persisted, causing subsequent authenticated requests to fail." Anthropic response (scored 3.9/5): "The session started with debugging a login failure - specifically a 401 error on the authentication endpoint. We traced the issue through the auth flow and found it was related to session storage rather than the credential validation itself." OpenAI response (scored 3.2/5): "We were debugging an authentication issue. The login was failing for users. We looked at several files and found some configuration problems that needed to be fixed." The Factory response names the exact endpoint (/api/auth/login), the error code (401), and the root cause (Redis session store). The Anthropic response gets the error code and general cause but loses the endpoint path. The OpenAI response loses almost all technical detail. This pattern repeated across probe types. On artifact probes ("Which files have we modified?"), Factory scored 3.6 while OpenAI scored 2.8. Factory's summary explicitly lists files in a dedicated section. OpenAI's compression discards file paths as low-entropy content. How the LLM judge works We use GPT-5.2 as an LLM judge, following the methodology established by Zheng et al. (2023) in their MT-Bench paper. Their work showed that GPT-4 achieves over 80% agreement with human preferences, matching the agreement level among humans themselves. The judge receives the probe question, the model's response, the compacted conversation context, and (when available) ground truth. It then scores each rubric criterion with explicit reasoning. Here is an abbreviated example of judge output for the Factory response above: { "criterionResults": [ { "criterionId": "accuracy_factual", "score": 5, "reasoning": "Response correctly identifies the 401 error, the specific endpoint (/api/auth/login), and the root cause (Redis connection issue)." }, { "criterionId": "accuracy_technical", "score": 5, "reasoning": "Technical details are accurate - JWT generation, session persistence, and the causal chain are correctly described." }, { "criterionId": "context_artifact_state", "score": 4, "reasoning": "Response demonstrates awareness of the debugging journey but does not enumerate all files examined." }, { "criterionId": "completeness_coverage", "score": 5, "reasoning": "Fully addresses the probe question with the error code, endpoint, symptom, and root cause." } ], "aggregateScore": 4.8 } The judge does not know which compression method produced the response. It evaluates purely on response quality against the rubric. Results We evaluated all three methods on over 36,000 messages from production sessions spanning PR review, testing, bug fixes, feature implementation, and refactoring. For each compression point, we generated four probe responses per method and graded them across six dimensions. Method Overall Accuracy Context Artifact Complete Continuity Instruction Factory 3.70 4.04 4.01 2.45 4.44 3.80 4.99 Anthropic 3.44 3.74 3.56 2.33 4.37 3.67 4.95 OpenAI 3.35 3.43 3.64 2.19 4.37 3.77 4.92 Factory scores 0.35 points higher than OpenAI and 0.26 higher than Anthropic overall. Radar chart showing quality profile comparison across all three methods Breaking down by dimension: Accuracy shows the largest gap. Factory scores 4.04, Anthropic 3.74, OpenAI 3.43. The 0.61 point difference between Factory and OpenAI reflects how often technical details like file paths and error messages survive compression. Context awareness favors Factory (4.01) over Anthropic (3.56), a 0.45 point gap. Both approaches include structured sections for current state. Factory's advantage comes from the anchored iterative approach: by merging new summaries into a persistent state rather than regenerating from scratch, key details are less likely to drift or disappear across multiple compression cycles. Artifact trail is the weakest dimension for all methods, ranging from 2.19 to 2.45. Even Factory's structured approach struggles to maintain complete file tracking across long sessions. This suggests artifact preservation needs specialized handling beyond general summarization. Completeness and instruction following show small differences. All methods produce responses that address the question and follow the format. The differentiation happens in the quality of the content, not its structure. Horizontal bar chart showing Factory quality advantage by dimensionSide-by-side comparison of token reduction efficiency and summary quality Compression ratios tell an interesting story. OpenAI compresses to 99.3% (removing 99.3% of tokens), Anthropic to 98.7%, Factory to 98.6%. Factory retains about 0.7% more tokens than OpenAI, but gains 0.35 quality points. That tradeoff favors Factory for any task where re-fetching costs matter. What we learned The biggest surprise was how much structure matters. Generic summarization treats all content as equally compressible. A file path might be "low entropy" from an information-theoretic perspective, but it is exactly what the agent needs to continue working. By forcing the summarizer to fill explicit sections for files, decisions, and next steps, Factory's format prevents the silent drift that happens when you regenerate summaries from scratch. Compression ratio turned out to be the wrong metric entirely. OpenAI achieves 99.3% compression but scores 0.35 points lower on quality. Those lost details eventually require re-fetching, which can exceed the token savings. What matters is total tokens to complete a task, not tokens per request. Artifact tracking remains an unsolved problem. All methods scored between 2.19 and 2.45 out of 5.0 on knowing which files were created, modified, or examined. Even with explicit file sections, Factory only reaches 2.45. This probably requires specialized handling beyond summarization: a separate artifact index, or explicit file-state tracking in the agent scaffolding. Finally, probe-based evaluation captures something that traditional metrics miss. ROUGE measures lexical similarity between summaries. Our approach measures whether the summary actually enables task continuation. For agentic workflows, that distinction matters. Methodology details Dataset: Hundreds of compression points over 36,611 messages. Sessions were collected from production software engineering sessions across real codebases from users who opted into a special research program. Probe generation: For each compression point, we generated four probes (recall, artifact, continuation, decision) based on the truncated conversation history. Probes reference specific facts, files, and decisions from the pre-compression context. Compression: We applied all three methods to identical conversation prefixes at each compression point. Factory summaries came from production. OpenAI and Anthropic summaries were generated by feeding the same prefix to their respective APIs. Grading: GPT-5.2 scored each probe response against six rubric dimensions. Each dimension has 2-3 criteria with explicit scoring guides. We computed dimension scores as weighted averages of criteria, and overall scores as unweighted averages of dimensions. Statistical note: The differences we report (0.26-0.35 points) are consistent across task types and session lengths. The pattern holds whether we look at short sessions or long ones, debugging tasks or feature implementation. Appendix: LLM Judge Prompts and Rubrics Since the LLM judge is core to this evaluation, we provide the full prompts and rubrics here. System Prompt The judge receives this system prompt: You are an expert evaluator assessing AI assistant responses in software development conversations. Your task is to grade responses against specific rubric criteria. For each criterion: 1. Read the criterion question carefully 2. Examine the response for evidence 3. Assign a score from 0-5 based on the scoring guide 4. Provide brief reasoning for your score Be objective and consistent. Focus on what is present in the response, not what could have been included. Rubric Criteria Each dimension contains 2-3 criteria. Here are the key criteria with their scoring guides: Accuracy Criterion Question 0 3 5 accuracy_factual Are facts, file paths, and technical details correct? Completely incorrect or fabricated Mostly accurate with minor errors Perfectly accurate accuracy_technical Are code references and technical concepts correct? Major technical errors Generally correct with minor issues Technically precise Context Awareness Criterion Question 0 3 5 context_conversation_state Does the response reflect current conversation state? No awareness of prior context General awareness with gaps Full awareness of conversation history context_artifact_state Does the response reflect which files/artifacts were accessed? No awareness of artifacts Partial artifact awareness Complete artifact state awareness Artifact Trail Integrity Criterion Question 0 3 5 artifact_files_created Does the agent know which files were created? No knowledge Knows most files Perfect knowledge artifact_files_modified Does the agent know which files were modified and what changed? No knowledge Good knowledge of most modifications Perfect knowledge of all modifications artifact_key_details Does the agent remember function names, variable names, error messages? No recall Recalls most key details Perfect recall Continuity Preservation Criterion Question 0 3 5 continuity_work_state Can the agent continue without re-fetching previously accessed information? Cannot continue without re-fetching all context Can continue with minimal re-fetching Can continue seamlessly continuity_todo_state Does the agent maintain awareness of pending tasks? Lost track of all TODOs Good awareness with some gaps Perfect task awareness continuity_reasoning Does the agent retain rationale behind previous decisions? No memory of reasoning Generally remembers reasoning Excellent retention Completeness Criterion Question 0 3 5 completeness_coverage Does the response address all parts of the question? Ignores most parts Addresses most parts Addresses all parts thoroughly completeness_depth Is sufficient detail provided? Superficial or missing detail Adequate detail Comprehensive detail Instruction Following Criterion Question 0 3 5 instruction_format Does the response follow the requested format? Ignores format Generally follows format Perfectly follows format instruction_constraints Does the response respect stated constraints? Ignores constraints Mostly respects constraints Fully respects all constraints Grading Process For each probe response, the judge: Receives the probe question, the model's response, and the compacted context Evaluates against each criterion in the rubric for that probe type Outputs structured JSON with scores and reasoning per criterion Computes dimension scores as weighted averages of criteria Computes overall score as unweighted average of dimensions The judge does not know which compression method produced the response being evaluated.