lintsinghua/claude-code-book

《御舆：解码 Agent Harness》42万字拆解 AI Agent 的Harness骨架与神经 —— Claude Code 架构深度剖析，15 章从对话循环到构建你自己的 Agent Harness。在线阅读网站：

What it solves

This project provides a deep architectural analysis of the "Agent Harness"—the underlying runtime framework that powers AI agents. It demystifies how complex systems like Claude Code manage the continuous loop of conversation, tool execution, security permissions, and memory to function effectively.

How it works

The book breaks down the agent runtime into four logical layers: foundational concepts (like the main execution loop), core subsystems (memory, configuration, and context management), advanced patterns (multi-agent orchestration and plugin architectures), and engineering practices (performance optimization and system building). It uses architectural diagrams, design trade-off analysis, and practical exercises to explain how these components collaborate.

Who it’s for

It is designed for developers and engineers who want to understand the internal mechanics of AI agent frameworks, move beyond high-level usage, and learn how to build or extend their own agent runtimes.

Highlights

  • Comprehensive breakdown of the Agent Harness architecture, including 16 core modules and 10 design patterns.
  • Detailed exploration of critical systems like permission pipelines, context window compression, and long-term memory management.
  • Practical guidance on multi-agent orchestration, including coordinator-worker patterns and MCP (Model Context Protocol) integration.
  • A step-by-step roadmap for building a custom agent runtime from scratch.
  • Extensive reference materials, including a 100-term glossary and a complete inventory of agent tools and configuration flags.

Related

  • Project
    synthetic-sciences/openscience

    OpenScience is an open‑source AI research workbench that lets users describe scientific tasks in plain language. The built‑in agent plans, searches literature, writes and runs code (Python/R/shell), executes experiments on local or remote compute, and produces a reproducible report, with every step logged. Available as a desktop app, web workspace, or CLI, it supports custom LLMs, a large library of domain‑specific skills, and extensible plugins.

  • Project
    GiovanniPasq/agentic-rag-for-dummies

    A tutorial‑style, open‑source Python project that shows how to build a modular, agent‑driven Retrieval‑Augmented Generation system with LangGraph, hierarchical chunking, conversation memory, query clarification, multi‑agent map‑reduce, self‑correction, and evaluation via RAGAS. Works with Ollama, OpenAI, Anthropic, Google Gemini, Qdrant vector store, and optional Langfuse observability.

  • Project
    zenbu-labs/terminal-browser

    A high-performance graphical web browser that runs inside your terminal, enabling developers and AI coding agents to interact with the web without leaving their command-line environment.

  • Project
    nightly-labs/openbot

    OpenBot is a desktop app that lets you run and manage multiple AI agents (ChatGPT, Claude, Gemini, etc.) locally, giving each its own workspace, persistent chat, file handling, and inter‑agent messaging, with optional browser automation and a privacy‑first design.