// the take
A free, Chinese-language book covering agentic AI end to end — planning, memory, tool use, multi-agent protocols, evals, and 'agentic coding' practices. It's aimed at engineers moving from calling an LLM API to building agents that plan and act, plus tech leads picking a framework. There's no actual agent code here; the Python in the repo is just the build tooling for rendering the book.
The table of contents is unusually complete for the space — it covers ReAct/Reflexion-style reasoning, vector and graph memory, MCP/A2A protocols, multi-agent architectures, and AgentOps (observability, failure modes, autonomy metrics) rather than just prompting tricks. The 'volatile facts' appendix is a smart idea: model/API/benchmark claims are isolated into one table with a stated 30-day recheck cadence instead of being scattered through static prose that silently goes stale. The build pipeline has real engineering behind it (CI, dependabot, typography and mermaid-rendering tests, automated PDF releases) — more rigor than most doc repos bother with.
Chinese only, no English edition, which halves the useful audience for a topic this global. Licensed CC BY-NC-SA, so you can't reuse any of it in commercial training material or internal docs without reworking it from scratch. It's a single-author survey spanning reasoning, multi-agent systems, enterprise governance, and AI safety/alignment — breadth over depth, no peer review, and no way to verify claims against the author's own practical experience running these systems at scale. Being a snapshot of a fast-moving field, most of chapters 8 (frameworks) and 11 (security/alignment) will likely be dated within 6-12 months despite the volatile-facts mitigation.
We read the README, the directory tree and a key source file in October 2026. Nobody pays to be listed. How we judge →