Li Bojie’s open textbook on AI agents has grown into a substantial engineering project: 10 chapters, 103 associated experiments and community translations across 13 languages. The AI Agent Book repository is being reorganized as version 2.0, with its manuscript, diagrams, code and build tooling published under the Apache 2.0 license.
The project also has unusual momentum. A GitHub Trending snapshot on August 18 recorded 556 stars added that day. That number signals developer interest, not independent proof that every lesson or experiment is sound. What makes the project worth examining is more concrete: it lets readers move from an architectural idea to runnable code, then inspect a public record of which ambitious demonstrations have actually cleared their stated evidence requirements.
That combination matters in a field crowded with agent tutorials that stop after a tool-calling demo. This book tries to cover the less glamorous engineering around context, memory, evaluation, post-training and failure analysis, while leaving much of the supporting material open to inspection.
A practical definition shapes the book
The manuscript starts from a compact formula: an agent is a large language model plus context plus tools. The English first chapter uses that abstraction to separate the model from the surrounding system that supplies instructions, history, external knowledge and actions. It is a useful editorial choice because it keeps the book from treating a stronger model as the answer to every engineering problem.
The 10-chapter route broadens steadily. It begins with agent basics, then moves through context engineering, user memory and knowledge bases, tool use, coding and general-purpose agents, richer interaction, evaluation, model post-training, continual improvement and multi-agent collaboration. The sequence makes clear that this is not just a prompt collection. It is an attempt to map the full stack around an agent, from a token window to a system that can act, be measured and change over time.
Version 2.0 is still being rearranged. The maintainers say chapters six through nine have been reordered around interaction, evaluation, post-training and continual evolution. They also warn that already-built PDFs can retain the older structure, so the repository README and chapter files are currently the authoritative index. Readers should treat 2.0 as an active manuscript rather than a frozen edition.
Chinese is the canonical text. English, Spanish, Indonesian, Arabic, Traditional Chinese, Russian, Tamil, Vietnamese, Japanese, Turkish, Korean and Hungarian editions are community contributions. That global reach is notable, but the repository explicitly says translations may lag behind the Chinese original. A translated chapter therefore should not be assumed to describe the latest code or numbering.
The experiments turn a syllabus into a repository
The headline figure of 103 experiments includes both projects implemented in the repository and external reproduction tracks. Their scope ranges from context compression, retrieval and memory policies to MCP tools, coding agents, computer control, speech systems, robotics, model training, benchmark runs and multi-agent coordination. Some are compact exercises; others require outside APIs, browsers, GPUs, emulators or physical hardware.
The basic setup is conventional Python, but it is organized by chapter. The project supports Python 3.11 through 3.13, exposes chapter-specific dependency groups and commits a uv.lock file for reproducible environments. A reader starting with chapter one can use:
git clone https://github.com/bojieli/ai-agent-book.git
cd ai-agent-book
uv sync --locked --extra ch1
uv run python chapter1/context/main.py
The repository also supports editable installation with pip, although that path resolves dependencies again instead of using the lock file. The distinction is important for a teaching project whose examples span fast-moving AI libraries. A working example today can break when a transitive package or model interface changes. A committed environment is not a guarantee, but it gives readers a defined starting point and makes bugs easier to compare.
There is no credible one-command promise for all 103 experiments. The CPU-friendly all dependency group excludes local training stacks, and individual projects can need CUDA, FFmpeg, browser binaries, Ollama or provider credentials. Later chapters also point to 22 external repositories plus a training cookbook, pinned to particular revisions for reproduction. Those checkouts bring their own requirements and licenses; Apache 2.0 on the book repository does not erase the terms attached to outside projects.
This is the right trade-off for breadth, provided readers notice it. The repository is best understood as a structured lab manual with several execution tracks, not as a single application waiting to be installed.
The evidence ledger is the more important release
The strongest part of the project may be its experiment status and evidence ledger. It defines “complete” as having substantive saved evidence for the manuscript’s stated gates. An installed dependency, a cloned upstream repository or a smoke test does not qualify. “Incomplete” means an implementation, execution or evidence requirement remains open, while some work is deliberately left as a reader-run exercise.
That sounds procedural, but it changes how the examples can be read. The ledger does not turn every run into a success story. In one active tool-discovery comparison, both control and discovery approaches completed three of three tasks. Discovery reduced exposed schema text and elapsed time, yet it did not improve accuracy. A model-distillation experiment likewise reports that a Qwen2.5 1.5B student rose only from one correct answer in 24 to two in 24, versus 23 for the teacher; the paired improvement was not statistically significant.
Other entries keep their scope deliberately narrow. An AndroidWorld run retains 580 episodes and reports just 26 strict successes. The record notes that its local Qwen model differs from the model used in the source it was paired with, so the result cannot establish a same-model improvement or even non-inferiority. That caveat is more useful than a broad claim built from incomparable setups.
The incomplete rows are equally revealing. Some tool integrations remain blocked by missing Calendar, Notion or email authorization. Robot tasks have source checkouts and non-actuating preflight work but no authorized physical run. A multilingual training exercise has an implementation but no retained checkpoint or before-and-after benchmark. These are ordinary constraints in agent research, and writing them down prevents repository presence from being confused with experimental completion.
The ledger is selective rather than a second index of all 103 exercises. Each chapter README remains the definitive list. Nor does retained evidence amount to peer review or independent reproduction. It does, however, give contributors and readers a place to challenge a bounded claim using the same manifests, receipts, hashes and acceptance criteria.
Useful, but not a neutral standard
The book is an authored curriculum, not an industry specification. Its formula and chapter order are design choices, and the field is moving fast enough that model names, APIs and recommended frameworks will age unevenly. The repository also includes a sponsor section promoting an API broker, something readers should distinguish from the technical evidence elsewhere in the project.
Those limits do not cancel the value of opening both the prose and the work behind it. Developers can inspect how an experiment was framed, run the smaller examples, swap a provider or model, and submit corrections. Educators can reuse diagrams and exercises within the license. Teams evaluating agent patterns can use the ledgers as examples of how to state a test boundary, preserve negative results and avoid treating a successful demo as a general capability.
The sensible entry point is not the most elaborate robot or multi-agent run. It is one chapter whose prerequisites match the reader’s environment, followed by the corresponding README and status record. That route makes the project’s real contribution easier to see: a common structure connecting explanation, implementation and evidence, even when the evidence says an idea did not work.
What to watch next is the stabilization of the 2.0 chapter structure, refreshed PDFs and translations, and progress on experiments still missing credentials, retained checkpoints or authorized hardware runs. Independent reproductions would carry more weight than another jump in stars. For now, the repository’s traction shows demand for a practical agent curriculum; its lasting value will depend on whether the code and evidence remain auditable as the manuscript changes.