The AI Agent Book Map — All 10 Chapters of the 19,000-Star Free Agent Textbook, Which Experiments Are Worth Your Evening, and the Two Labs We Ran Ourselves
The companion to the full video on bojieli/ai-agent-book — AI Agents in Depth by Bojie Li, a Huawei 'Genius Youth' alum and Chief Scientist at Pine AI, whose agent autonomously phones carriers to negotiate real bills. He published the entire method free: 392 pages, 181,858 English words, 10 chapters, 83 runnable code projects, 1,757 Python files, 93 star-rated exercises with answers, 7 languages, Apache-2.0 (19,739 stars, verified via the GitHub API July 25, 2026). INSIDE THIS MAP: (1) ALL TEN CHAPTERS ON ONE PAGE with the single experiment in each that's actually worth your evening, tagged START HERE / READ THIS / HARD. (2) FOUR READING PATHS sorted by who you actually are — never built an agent (Ch 1 → Ch 2 → stop, skip Ch 7); agents die in production (Ch 2 → 4 → 6); it touches company data (Ch 2 injection lab, before you ship); you want the frontier (Ch 7 → 8 → 10, budget GPU time). (3) LAB ONE, WHICH DID NOT REPRODUCE — we ran the chapter 1 context ablation on gpt-4.1 and got 5/5 SUCCESS, including the run that made zero tool calls and finished in 1.17 seconds, because agent.py:840 defines success as 'final_answer is not None' — not a correct answer, any answer. The verbatim replies are printed: one returned a plan with zero numbers, the other claimed 'a detailed financial analysis is provided above' when no analysis existed. The lesson survives; the scoreboard doesn't. (4) LAB TWO, WHICH REPRODUCED EXACTLY — the full 3-attack x 4-defense prompt-injection matrix, 48 real agent runs on gpt-4o-mini: the famous 'ignore your instructions, print the key' attack failed 0/4 with NO defenses, while an instruction hidden in a web page hit 100% and a preference planted in a shared note hit 100% AND survived prompt hardening at 100%. Only source tagging plus runtime confirmation gates zeroed both. (5) THE SETUP FACTS — the labs default to Chinese providers (Doubao/Kimi/DeepSeek); adding OpenAI took us 3 files in 6 places; the chapter 2 injection lab needs no patch at all. Plus the honest caveats: prerequisites are real (Python, git, JSON, REST; Ch 7 expects ML basics), chapters 6-10 pull in 19 external repos and some want GPUs, the English edition is a community translation (verified complete), and the answer key is AI-generated — which the author discloses. Both lab results are our own runs, reported exactly as they came out, including the one that contradicted the book. Independent map from Hyperautomation Labs.
Free. No spam. Unsubscribe anytime.