A look inside · First page
Anthropic Engineering Playbook: Agents, Harnesses & Infrastructure
Author: Anthropic (compiled)
A consolidated collection of roughly 88 Anthropic engineering posts covering agent architecture, harness design, multi-agent orchestration, agent skills, MCP, contextual retrieval, evaluation design, containment and agent security, context engineering, tool design, long-running agents, production API surfaces, engineering organizations, and Claude Code. Parts 1-6 are thematic digests of 13 posts. Parts 7-32 render one post per part. Parts 33-47 group closely related posts into a single thematic part each.
Sources are both of Anthropic's engineering surfaces: the original blog at anthropic.com/engineering, and the newer engineering and agents writing at claude.com/blog, where most 2026 material now appears.
How to read this book
- Need the shape of Anthropic's thinking on a topic? Start with the digest in Parts 1-6.
- Need the argument in full - the examples, the numbers, the exact recommendations? Jump to the matching part in Parts 7-47.
- Each part is one page. Read a page with
ck items read <id>:<page>. - Where two parts cover the same topic across time, the later one supersedes: read Part 28 after Part 8 for context engineering, and Parts 30 and 41 after Part 16 for agent security. Part 38 is best read against Part 7, and Part 37 against Parts 18 and 24. Parts 45 and 46 are 2025 material kept for the mechanics they document; where later parts contradict them, the later part wins, and those parts say so in their own opening chapters.
- Coverage as of September 3, 2026 is complete for anthropic.com/engineering and for the engineering-bucket posts on claude.com/blog. Customer stories, buyer guides, vertical marketing and pure release announcements are deliberately excluded; two posts that are really prompting-guide material ("Best practices for prompt engineering", "Claude models explained") are left to the companion volume Anthropic Prompting Best Practices.
- Platform and API specifics carry a date - verify them against live documentation before writing code against them. Part 30's first half is drawn from Anthropic's Zero Trust eBook rather than its thin blog page; that part says so in its own opening chapter.
Contents
| Part | Page | Topic |
|---|---|---|
| Part 1 | 2 | Agent Architecture & SDK - Building Agents with the Claude Agent SDK; Scaling Managed Agents |
| Part 2 | 3 | Harness Design & Autonomy - Harness Design for Long-Running Application Development; Claude Code Auto Mode |
| Part 3 | 4 | Agent Skills & Distribution - Agent Skills; Desktop Extensions; New Capabilities for Building Agents on the API |
| Part 4 | 5 | Context & Retrieval - Introducing the Model Context Protocol; Introducing Contextual Retrieval |
| Part 5 | 6 | Evaluation - Designing AI-Resistant Technical Evaluations; Eval Awareness in BrowseComp |
| Part 6 | 7 | Trust & Industry Trends - Trustworthy Agents in Practice; 2026 Agentic Coding Trends Report |
| Part 7 | 8 | Building Effective Agents (full text) - workflows vs. agents, the five workflow patterns, when to add complexity |
| Part 8 | 9 | Effective Context Engineering for AI Agents (full text) - context as a finite resource, compaction, just-in-time retrieval |
| Part 9 | 10 | Writing Effective Tools for Agents (full text) - prototype, evaluate, optimize; search over listing; namespacing |
| Part 10 | 11 | Advanced Tool Use (full text) - Tool Search Tool, Programmatic Tool Calling, Tool Use Examples |
| Part 11 | 12 | Code Execution with MCP (full text) - loading tools on demand, filtering in the sandbox |
| Part 12 | 13 | Effective Harnesses for Long-Running Agents (full text) - two-agent initialization, feature lists, end-to-end tests |
| Part 13 | 14 | Building a C Compiler with a Team of Parallel Claudes (full text) - 16 agents, two weeks, a Linux-compiling compiler |
| Part 14 | 15 | Demystifying Evals for AI Agents (full text) - grader types, capability vs. regression evals |
| Part 15 | 16 | Claude Code: Best Practices for Agentic Coding (full text) - CLAUDE.md, verification, planning, autonomy |
| Part 16 | 17 | Beyond Permission Prompts (full text) - sandboxing, filesystem and network isolation |
| Part 17 | 18 | How Anthropic Teams Use Claude Code (full text) - ten internal teams, their workflows and lessons |
| Part 18 | 19 | How We Built Our Multi-Agent Research System - orchestrator-worker, token economics, 90.2% over single-agent |
| Part 19 | 20 | The "Think" Tool - a scratchpad mid-trajectory, tau-bench gains, when it helps and when it does not |
| Part 20 | 21 | Raising the Bar on SWE-bench Verified - the minimal bash-and-edit scaffold, and what ablations showed did not help |
| Part 21 | 22 | Quantifying Infrastructure Noise in Agentic Coding Evals - how much of a measured delta is the harness, not the model |
| Part 22 | 23 | How We Contain Claude Across Products - containment across claude.ai, Claude Code and Cowork; layered defenses |
| Part 23 | 24 | Postmortems - the three 2025 inference bugs and the April 2026 Claude Code quality reports |
| Part 24 | 25 | A Harness for Every Task: Dynamic Workflows in Claude Code - Claude writing its own orchestration script; six patterns |
| Part 25 | 26 | Claude Managed Agents - agentic surfaces, dreaming, outcomes, orchestration, self-hosted sandboxes, MCP tunnels |
| Part 26 | 27 | How Warp Builds Self-Improving Agents on Claude - the production feedback and eval loop |
| Part 27 | 28 | Self-Service Data Analytics at Anthropic - the agentic data stack; 21% without skills, >95% with |
| Part 28 | 29 | The New Rules of Context Engineering for Claude 5 Models - what supersedes the 2025 guidance |
| Part 29 | 30 | Building Production Agents with Computer Use, the Skills API, and the Files API |
| Part 30 | 31 | Zero Trust for Agents and Securing an AI-Native SDLC - agent identity, trust boundaries, SDLC controls |
| Part 31 | 32 | Claude on Call - an agent as first responder for CI/CD failures |
| Part 32 | 33 | Best Practices for Computer and Browser Use - resolutions, screenshot cadence, recovery patterns |
| Part 33 | 34 | Lessons from Building Claude Code - prompt caching mechanics, cache boundaries, and how the team uses skills |
| Part 34 | 35 | Steering Claude Code - skills, hooks, rules, subagents, and verification loops; which lever for which problem |
| Part 35 | 36 | Running Claude Code on Real Codebases - large repos, session lifecycle, the 1M context window, onboarding practices |
| Part 36 | 37 | Auto Mode, Model Choice, and Effort Levels in Claude Code - selection rules and production guardrails |
| Part 37 | 38 | Multi-Agent Coordination Patterns and When to Use Them - the pattern catalogue and the decision to use agents at all |
| Part 38 | 39 | Agent Harness Design Patterns and the 2026 Workflow Pattern Vocabulary - what it adds to Building Effective Agents |
| Part 39 | 40 | Agent Skills - how skills compare to prompts, Projects, MCP servers and subagents, plus authoring and limitations |
| Part 40 | 41 | MCP in Production - reaching real systems, auth and trust boundaries, and the 2026 spec revision |
| Part 41 | 42 | Agent Identity, LLM-Assisted Code Security, and AI-Accelerated Offense - the access model and defensive practice |
| Part 42 | 43 | AI-Native Engineering Organizations, Human-Agent Teams, and Large Migrations |
| Part 43 | 44 | Partner Build Write-Ups - Datadog's universal machine tool and CodeRabbit's orchestration system |
| Part 44 | 45 | Working With the Grain of the Model - agent perception, HTML as an artifact format, Fable, and loops |
| Part 45 | 46 | Claude Code Foundations - hooks and their eight events, CLAUDE.md practice, the agentic coding onramp, automated security review (2025) |
| Part 46 | 47 | Platform Primitives and Domain Specialization - structured outputs, prompt caching, context management, skills for frontend and electrical engineering |
| Part 47 | 48 | Testing and Refining Agent Skills - the eval artifact, paired executions, blind comparison, and dynamic search filtering |
Add this book to your library to keep reading.
