Open-Source AI Agent Book: 10 Chapters, 103 Runnable Experiments

The article introduces an open-source book titled Deep Understanding of AI Agent: Design Principles and Engineering Practice (《深入理解 AI Agent：设计原理与工程实践》) authored by Li Bojie, an engineer with years of AI Agent experience. The book is hosted on GitHub at bojieli/ai-agent-book with 39k+ stars, written in Python, and licensed under Apache-2.0.

Core Formula: Agent = LLM + Context + Tools

The entire book revolves around a single formula: Agent = LLM + Context + Tools . The author argues this decomposition captures the essence of an Agent: the LLM handles understanding and decision-making, context supplies background and memory, and tools turn decisions into actions. Missing any component breaks the Agent.

Four-Part Structure Across Ten Chapters

Foundations (Ch.1) : Defines Agent and establishes that engineering competitiveness lies in the Harness (outer framework), not just the model.

Capabilities (Ch.2–4) : Covers context engineering (KV Cache, prompt engineering, context compression), user memory and knowledge bases (RAG, structured indexing, knowledge graphs), and tools (MCP protocol, perception/execution/collaboration tool categories, proactive tool discovery).

Extension (Ch.5–8) : Explores multimodal interaction, evaluation (building comparable test signals), and model post-training (when to use SFT vs. RL, tool-call internalization).

Evolution & Collaboration (Ch.9–10) : Addresses continuous learning from execution traces, knowledge updates, and multi-agent collaboration demonstrating emergent group intelligence.

103 Runnable Experiments

The book distinguishes itself by providing 103 runnable experiments — one per major concept — so readers can verify claims hands-on. Experiments use Python 3.11–3.13 and are managed with uv. For example, to run Chapter 1 experiments:

uv sync --locked --extra ch1
uv run python chapter1/context/main.py

API-key-required experiments include setup guides for Kimi, Zhipu, DeepSeek, and OpenRouter.

Three Ways to Use the Book

Read directly via the online site ( bojieli.github.io/ai-agent-book) or download PDF/EPUB.

Run experiments — the core value — by cloning the repo and executing chapter-specific code.

Contribute via pull requests: fix text, bugs, add experiments, or translate (already 14 languages).

Comparison with Other Resources

vs. scattered tutorials : Systematic, one formula through ten chapters, logical continuity.

vs. pure theory books : Adds 103 runnable experiments for empirical verification.

vs. framework-specific tutorials : Teaches universal principles and engineering methodology, not tied to a single framework.

The book assumes basic Python and LLM knowledge; absolute beginners should start elsewhere.

Key Takeaways

Decompose complex systems into independently researchable components (LLM, Context, Tools).

Context engineering determines the capability ceiling — better to optimize context than chase stronger models.

Real engineering differentiation lies in the Harness: memory, tools, evaluation, training.

The article concludes that for anyone serious about mastering AI Agents, this open-source, experiment-driven book is a rare systematic resource worth studying and running.

GitHub: github.com/bojieli/ai-agent-book