Open Source
Open Source

ai-agent-book: Open-Source Systematic AI Agent Learning (10 Chapters, 95 Experiments, 13 Languages, 33K Stars)

bojieli/ai-agent-book (33,213 stars, Python, Apache-2.0, created 2025-09-09, pushing today, GitHub Trending) is the open-source repo for the in-depth AI Agent book by Li Bojie: Design Principles and Engineering Practice. Agent = LLM + context + tools; 10 chapters from principles to engineering plus 95 hands-on experiments (local + external repro) plus 13 languages; PDF/EPUB free. The top pick for developers who want to learn AI Agent systematically; the moat is systematic depth plus experiments plus multilingual plus open and free.

Published August 6, 20268 min read
<!-- ai-agent-book-resource | open-source | ai-agent-book: An Open-Source Systematic AI Agent Learning Resource (10 Chapters, 95 Experiments, 13 Languages) -->

AI Agent has blown up this year, but anyone trying to learn it systematically hits an awkward gap: what is available is either scattered vendor-blog tutorials or theoretical paper frameworks. The middle layer, something that runs all the way from principles to engineering practice and lets you run things by hand, has been missing. Through the second half of 2025 the Agent framework war mostly settled, and components like tooling, memory, and evaluation each found de facto standards, but the one book a developer could pick up, read end-to-end, and know how to build afterward had still not been written. bojieli/ai-agent-book fills that gap: an open-source AI Agent book, 10 chapters of main text with 95 experiments, free downloads in 13 languages, still pushing today.

What It Is

ai-agent-book (github.com/bojieli/ai-agent-book) is an open-source project by Li Bojie, corresponding to the book "In-Depth Understanding of AI Agent: Design Principles and Engineering Practice". It has 33,213 GitHub stars, 3,565 forks, primary language Python, Apache-2.0 license, created on 2025-09-09, with its latest push today (2026-08-06). It has been a GitHub Trending Project of the Day and currently sits at #2 on the OpenGithubs weekly rank.

The whole book is built around one core formula: Agent = LLM + Context + Tools. The line looks simple, but the author breaks it into 10 progressively deeper chapters, starting from the judgment that "Harness engineering is the real competitiveness" and walking through context, memory, tools, coding agents, evaluation, post-training, continuous evolution, multimodality, and multi-agent collaboration. What separates it from scattered tutorials is "book-level coherence": a complete narrative line, each chapter backed by runnable experiments, not a slide deck mashup.

Core Capabilities

Three pillars: 10 chapters of main text, 95 companion experiments, and 13 languages. The chapter skeleton (per-chapter project counts come from the repo README):

ChTopicOne-line coreProjects
1Agent fundamentalsAgent = LLM + Context + Tools; Harness engineering is the real competitiveness4
2Context engineeringContext sets the capability ceiling: KV Cache, prompt engineering, Agent Skills, context compression9
3User memory and knowledge basesRemember users across sessions, plug in external knowledge: user memory, RAG, structured indexing, knowledge graphs13
4ToolsTools are the agent's hands: MCP protocol, perception/action/collaboration tool types, event-driven async agents, proactive tool discovery7
5Coding agents and code generationCode is "a tool that can create new tools"; a production-grade coding agent landscape12
6Agent evaluationTurn performance into comparable signals: eval environments, metrics, statistical significance, eval-driven selection12
7Model post-trainingPretraining / SFT / RL in three stages: when to choose SFT, when to choose RL, tool-call internalization, sample efficiency16
8Continuous evolution of agentsExtract learning signals from trajectories, update knowledge, instructions, programs, and parameters9
9Multimodality and real-time interactionExtend from text to voice, GUI, the physical world: three voice paradigms, Computer Use, robotics10
10Multi-agent collaborationCollective intelligence above the individual: collaboration frameworks, context sharing/isolation, an emergent "agent society"8

The 95 experiments run on two tracks: local projects (built into the repo, installable via uv sync) and external reproduction tracks (22 external repos + 1 helper cookbook mapped to chapters 6, 7, 9, and 10, which you clone into the corresponding directories yourself). Each experiment carries one of three status marks: ✅ runnable (config the environment and it produces the expected result), 📖 reproduction (reproduces a published paper or project result), and 🚧 design (gives the design idea and code skeleton, not yet fully implemented), so you can see at a glance how far it can go. Execution status, evidence, and unfinished gates are tracked separately in docs/EXPERIMENT_STATUS.md; cloning the repo does not mean the experiments are done, the README is explicit about this.

The 13 languages: Chinese (original), English, Spanish, Indonesian, Arabic, Traditional Chinese (Taiwan), Russian, Tamil, Vietnamese, Japanese, Turkish, Korean, Hungarian. Everything except the Chinese original is a community translation.

How to Use

Three paths, ordered by barrier to entry:

  1. Download PDF / EPUB directly (least hassle). The repo Release always points to the latest build on the main branch; both the Chinese original and the 12 community translations have PDF + EPUB, open-source and free. Fixed versions live on the Releases page.
  2. Read online: bojieli.github.io/ai-agent-book supports multilingual switching, chapter folding, full-text search, and direct links to companion experiments, automatically rebuilt after every push to main.
  3. Compile it yourself: to edit content or produce a local PDF, install pandoc, xelatex, the ElegantBook document class, and the relevant fonts, then run bash build_pdf.sh inside book/. Source lives in book/introduction.md, book/chapter1.md through chapter10.md, and book/afterword.md; figures in book/images/; layout details in preamble.tex and *.lua.

To run experiments: Python 3.10+ is required, uv is recommended (uv sync --locked --extra ch1, swapping ch1 for ch2 through ch10 to install a given chapter), and pip works too (python -m pip install -e ".[ch1]"). Experiments that call models need you to configure an API key (copy .env.example to .env and fill in at least one provider key; some experiments want a .env in their own directory or an exported env var).

Pitfalls

  1. Community translations lag the Chinese original. The README states explicitly that the 12 non-Chinese editions (English, Spanish, Indonesian, Arabic, Traditional Chinese, Russian, Tamil, Vietnamese, Japanese, Turkish, Korean, Hungarian) are community contributions and "may lag behind the Chinese original." For the latest content, read Chinese; treat the others as backups.
  2. The build environment has a threshold. The pandoc + xelatex + ElegantBook + Chinese-fonts combo can stall someone who has never touched LaTeX for half a day. If you only want to read, do not compile; just download the PDF.
  3. External reproduction tracks depend on third-party environments. The 22 external repos (AndroidWorld, GAIA, OSWorld, SWE-bench, tau2-bench, terminal-bench, MiniMind, verl, SandboxFusion, claude-quickstarts, and others) are not built-in dependencies; you have to clone them into the right directories yourself, and some training projects also pull models, datasets, and simulators per their own READMEs. The hardware bar is not low.
  4. Python version fragmentation. The baseline is 3.10+, but some browser and memory experiments need 3.11+, and some built-in third-party components in chapter 8 need 3.12+. A single environment will not run every experiment; you have to configure per chapter.
  5. Experiments ≠ completed. EXPERIMENT_STATUS.md tracks execution status separately; pinned source only establishes a reproduction starting point and does not mean training, hardware, browser, or multi-agent experiments have been run. Do not assume cloning is the finish line.

Take

ai-agent-book hits a genuine void: AI Agent has lacked a book you can read end-to-end and also run by hand. Its moat is not any single chapter being deep, but the four-pack of "systematic main text + runnable experiments + multilingual + open-source free": 10 chapters complete the narrative from principles to engineering, 95 experiments let readers touch instead of just read, 13 languages push the barrier down globally, and Apache-2.0 plus free PDF downloads let it spread. The line in chapter 1, "Harness engineering is the real competitiveness," is the eye of the book; the author is not teaching how to use some framework but laying out the full landscape of agent engineering.

It fits two crowds: developers who want to learn AI Agent systematically and are not satisfied with scattered tutorials, and researchers working on agents who need a reproducible starting point for experiments. The barrier is that you need a Python environment, willingness to configure API keys, and, if you actually want to run the chapter 7 training experiments, GPUs. Against vendor docs, its edge is that it is not tied to one framework and is not selling anything: chapter 4 on tools does not just push the author's own MCP implementation, and chapter 6 on evaluation lines up the mainstream benchmarks side by side. Against papers, its edge is that every chapter lands in a runnable experiment instead of stopping at concepts. If you only want a popular-science intro to agents, this book is a bit hard; but if the one you want is "can land, can reproduce, can ask why," 33,213 stars and today's push say it is the one.


References

This article is AI-assisted and human-edited. Last updated: 2026-08-06

Related

Open Source

Sequence In, Structure Out: Why Did a 62-Star Repo Make Nature?

structRFM is a structure-guided RNA foundation model from the MIRACLE lab at USTC; the paper appeared in Nature on November 5, 2026 (DOI 10.1038/s41586-026-11210-2). The repository heqin-zhu/structRFM is MIT licensed and showed 62 stars in a November 6, 2026 GitHub API check; it was created in February 2025 and its README still points to the August 2025 bioRxiv preprint. The team says the roughly 21 million sequence-structure pairs, model weights and source code are public: the dataset is built from RNAcentral and BPfold with a 512-token length cap and is 4.5 GB on Zenodo. SgMLM structure-guided masked language modeling and MUSES multi-source structure ensembling let the model emit sequence-level, nucleotide-level and pair-level features from sequence alone, supporting secondary and tertiary structure prediction, splice site prediction, IRES identification, ncRNA classification and zero-shot homology classification.

Nov 6, 20269 min read
Open Source

17 Seconds on a CPU: Can an API Prove Which Model Answered?

anthropics/FlashProve is an open-source research prototype for zero-knowledge verifiable inference from Anthropic (Apache-2.0, 843 stars per a November 5, 2026 GitHub API check, version 0.16.0, officially labeled a research artifact: not maintained and not accepting contributions). The three-step protocol: the provider publishes a weight commitment ID offline (about 39 minutes on one H100), attaches a GPU-generated proof to each reply (about 19MB per 1024 tokens), and anyone verifies with cp_verify on a plain CPU in seconds. It runs Llama 4 Maverick at full 400B size on a single H100 at about 110 tokens per second of proof prefill with about 17-second CPU verification, at a cost of roughly 70 to 95 times normal inference GPU time; soundness is formally proved in Lean 4. It proves provenance, not correctness, and no major provider ships proofs yet.

Nov 5, 20269 min read
Open Source

Even Failed Experiments Trained It: An MIT Science LLM

On November 2, Deep Principle released and open-sourced Hydrogen-1.0: a 753B science LLM post-trained from GLM-5.3, under MIT, live on ModelScope and HuggingFace (deep-principle/Hydrogen-1.0, about 1.51TB BF16, plus NVFP4/MXFP4 variants, tested on 8 GPUs with SGLang). Vendor-reported FSR 38.3% beats GPT-6-Astra at 33.3% and GPQA-Diamonds 93.0% nears Kimi-K3; its signature is training on its own lab records including failed experiments, with a Lab Compile step that catches faulty experiment instructions.

Nov 4, 20269 min read