AI Agent Book

This repository holds the open-source book AI Agents in Depth: Design Principles and Engineering Practice by Li Bojie, along with its full text, compiled PDF/EPUB editions, and 94 companion code experiments organized by chapter. It's aimed at engineers and researchers who want to understand how AI agents actually work under the hood, not just how to call an agent framework's API. The book frames agents around a simple formula: Agent = LLM + context + tools, and builds out ten chapters from that base, ending in production-grade topics like evaluation, post-training, and multi-agent coordination.

The content is written in Chinese, with community translations into English, Arabic, Traditional Chinese (Taiwan), Russian, Tamil, Vietnamese, Japanese, Turkish, and Korean, though translations may lag behind the original. Each chapter links to runnable code, reproduction tracks, or design sketches, so readers can pair concept with implementation instead of just reading theory.

Key features

• Ten-chapter structure: covers agent basics, context engineering, memory and RAG, tool use (MCP), coding agents, evaluation, post-training, continual learning, multimodal/real-time interaction, and multi-agent collaboration.
• 94 companion experiments: labeled as runnable, reproduction, or design-stage projects, spread across the chapters with per-chapter README files.
• Multi-format distribution: read chapters directly on GitHub in Markdown, download compiled PDF/EPUB, or use the online reader with search and multi-language switching.
• Ten-language coverage: original Chinese text plus nine community-maintained translations, each in its own book-<lang>/ directory.
• External repo mapping: a clone script pins 20 external repositories (benchmarks like SWE-bench, GAIA, OSWorld, training frameworks like verl, and hardware/robotics tracks) to specific commits for reproducibility.// no changed// keep list under 10 items
• Apache-2.0 license: both the book text and code are open for reuse and modification.

Ideal use cases

This repo fits people studying agent architecture who want a structured, code-backed curriculum: engineers building coding agents, RAG pipelines, or MCP-based tool integrations; researchers evaluating agent benchmarks (GAIA, SWE-bench, OSWorld); and teams designing multi-agent systems who want reference implementations rather than blog-post summaries. It's also useful as a syllabus for a study group, since chapters map to concrete, gradable experiments.

It's not a good fit if you're looking for a ready-made agent framework or SDK to drop into production. The code here is instructional and experiment-scoped, not a maintained library with stable APIs. Some chapters (7, 9, 10) depend on external repos pinned to specific commits, hardware (robots), or paid API keys, so a fully offline, zero-dependency read-through of every experiment isn't realistic. If your goal is just