An open-source AI agent textbook where every chapter runs

What it does

This is the master repository for 《深入理解 AI Agent：设计原理与工程实践》, a Chinese technical book by Li Bojie on designing and building AI agents. It hosts the complete book text in Markdown, a compiled PDF, and runnable Python projects organized across ten chapters. The material treats agent building as a systematic engineering discipline rather than folklore.

The interesting bit

Most AI books give you prose or isolated snippets; this one treats every concept—from KV Cache layout to MCP servers to GraphRAG—as a standalone, executable project. The author frames the entire field around the formula Agent = LLM + Context + Tools, then systematically builds out each term with working code.

Key highlights

• Ten chapters span the full stack: context engineering, memory/RAG, tool use (MCP), coding agents, evaluation, RL/post-training, self-evolution, multimodal interaction, and multi-agent collaboration.
• Each chapter contains independent, runnable Python examples: attention visualization, BM25 sparse retrieval built from scratch, Agentic RAG with ReAct loops, and a local-LLM serving demo showing that even a 0.6B model can handle tool calls with the right system design.
• Includes advanced indexing implementations: RAPTOR recursive trees, GraphRAG knowledge graphs, and Anthropic-style contextual retrieval (reducing retrieval failure rates by 49–67% in the included demo).
• Book source is fully open: Markdown files, build scripts for PDF generation (via pandoc/XeLaTeX), and figure-generation scripts.
• Some demos rely on free public APIs (DuckDuckGo, Open-Meteo) to lower the barrier to entry.

Caveats

• The book text and code comments are in Chinese, which limits accessibility for non-Chinese readers.
• PDF compilation requires a specific LaTeX toolchain (pandoc, XeLaTeX, the ElegantBook document class, and Chinese fonts) that may not be trivial to set up.
• Several demos require API keys for commercial LLM providers, though some chapters use free public APIs.

Verdict

Worth bookmarking if you want a structured, code-first curriculum covering the modern agent stack end-to-end. Skip it if you are looking for a single deployable framework or need English-language documentation.