54.7K ·GitHub
What is this?
An open-source book by Li Bojie taking AI agents from design principles to engineering practice, built on the formula Agent = LLM + context + tools; text, figures and labs all live in this repo.
An open-source book by Li Bojie taking AI agents from design principles to engineering practice, built on the formula Agent = LLM + context + tools; text, figures and labs all live in this repo.Is this a book you just read, or code you run?Both: the book has 10 chapters you can read as markdown on GitHub, as PDF/EPUB downloads, or in a web reader with full-text search, and it ships 109 runnable companion labs.
Both: the book has 10 chapters you can read as markdown on GitHub, as PDF/EPUB downloads, or in a web reader with full-text search, and it ships 109 runnable companion labs.How do I get the companion labs running?Install uv, then from the repo root run uv sync --locked --extra chN for your chapter (ch1 through ch10); pip install -e '.[chN]' also works but skips the lockfile.
Install uv, then from the repo root run uv sync --locked --extra chN for your chapter (ch1 through ch10); pip install -e '.[chN]' also works but skips the lockfile.Besides Python, what does running a lab require?A model provider key: copy .env.example to .env at the repo root and fill in at least one provider's key; some labs want it in their own directory or as exported env vars.
A model provider key: copy .env.example to .env at the repo root and fill in at least one provider's key; some labs want it in their own directory or as exported env vars.Will every experiment run on a plain laptop?Not all of them. The 'all' extra is a CPU-friendly set without the local training stack; some chapter 8 components need Python 3.12+, and heavier runs need CUDA per the lab README.
Not all of them. The 'all' extra is a CPU-friendly set without the local training stack; some chapter 8 components need Python 3.12+, and heavier runs need CUDA per the lab README.Does any of this depend on my operating system?Reading is OS-independent. The labs target Python 3.11-3.13 and run on any OS; system extras like browsers, CUDA, FFmpeg, Ollama and Playwright are documented in each lab's README.
Reading is OS-independent. The labs target Python 3.11-3.13 and run on any OS; system extras like browsers, CUDA, FFmpeg, Ollama and Playwright are documented in each lab's README.What do the ten chapters actually cover?It walks from the core formula through context engineering (KV cache, prompts, skills, compression), user memory and RAG, tools and MCP, coding agents, multimodal interaction, evaluation, post-training (SFT vs RL), continual evolution, and multi-agent collaboration.
It walks from the core formula through context engineering (KV cache, prompts, skills, compression), user memory and RAG, tools and MCP, coding agents, multimodal interaction, evaluation, post-training (SFT vs RL), continual evolution, and multi-agent collaboration.Can I read it in English or other languages?Yes — 15 languages in total. English, Spanish, Indonesian, Russian, Japanese, Korean and more are community translations (book-en/ etc.) that may lag behind the Chinese original.
Yes — 15 languages in total. English, Spanish, Indonesian, Russian, Japanese, Korean and more are community translations (book-en/ etc.) that may lag behind the Chinese original.Why does the README pin 22 external repos like SWE-bench and OSWorld?Chapters 6, 7, 8 and 10 map experiments to 22 external repos plus a training cookbook. For size and copyright reasons they are not bundled; a clone script checks out pinned 2026-07-30 commits as reproduction starting points.
Chapters 6, 7, 8 and 10 map experiments to 22 external repos plus a training cookbook. For size and copyright reasons they are not bundled; a clone script checks out pinned 2026-07-30 commits as reproduction starting points.How do I know which experiments actually run end to end?docs/EXPERIMENTSTATUS.md tracks each lab's execution status, evidence and unfinished gates — cloning or installing the source does not mean an experiment has been completed.
docs/EXPERIMENTSTATUS.md tracks each lab's execution status, evidence and unfinished gates — cloning or installing the source does not mean an experiment has been completed.What are similar projects?ai-system-design, autumn-recruitment, zhijian-ai-bluebook-workbuddy-harness, hello-agents, aie-book, Agentic-Design-Patterns, ai-agents-the-definitive-guide, llm_interview_note
ai-system-design, autumn-recruitment, zhijian-ai-bluebook-workbuddy-harness, hello-agents, aie-book, Agentic-Design-Patterns, ai-agents-the-definitive-guide, llm_interview_noteWhich topics is it in?experiments, epub, computer-use, llm-agents, ebook, robotics, knowledge-graph, context-engineering, reinforcement-learning, evaluation, agent-memory, fine-tuning
experiments, epub, computer-use, llm-agents, ebook, robotics, knowledge-graph, context-engineering, reinforcement-learning, evaluation, agent-memory, fine-tuning