Agentic AI Guide ZH: the 598-page Hitchhiker's Guide, translated into Chinese

A translation that arrived fast, and says so

The Hitchhiker's Guide to Agentic AI, a 598-page, thirty-chapter tour from Transformer internals to deployed agent systems, was written by Haggai Roitman and released under CC BY-SA 4.0. This repository is an unofficial Chinese translation of that guide, published as a derivative work under the same licence by the translator Chasing1020, and the README is upfront about three things at once: it is not endorsed or reviewed by the original author, the translation used large language models as an aid with human review on top, and any translation errors belong to the translator, with hallucinations named as the expected failure mode.

For Chinese-reading engineers, the value proposition is time. The original is a substantial English reference; this edition exists so readers can build the full picture, from hardware to training algorithms to agent systems, in their own language, with the terminology conventions handled systematically rather than improvised.

What the book covers: six parts, thirty chapters

The structure follows the original. Part one lays foundations: how a Transformer processes text, the GPU hardware and distributed systems that make training possible, and the first principles of reinforcement learning, answering questions such as why training a 70B model takes at least an eight-GPU node. Part two is the largest of the training half, covering PPO, DPO, GRPO and their many variants, reward model training, supervised fine-tuning practice, large-scale RLHF system architecture and trajectory-level RL for agents, each algorithm with derivations, intuition and runnable templates for the TRL library.

Part three covers reasoning, where reinforcement learning makes chain-of-thought, self-verification and backtracking emerge, plus test-time scaling and the models associated with it. Part four modernizes evaluation from classical metrics to LLM-as-judge, preference dimensions, ELO ranking, agent benchmarks and contamination detection. Part five, chapters 15 through 27, is the agent half: knowledge with RAG, memory, runtime orchestration, collaboration protocols such as MCP and A2A, multi-agent systems and agent UI, with comparisons of mainstream frameworks. Part six closes with 108 self-test questions with detailed solutions, a quick-reference chapter and an outlook.

Building the book yourself

The repository ships LaTeX sources, and the build requires XeLaTeX. The Makefile drives it:

make distclean && make

On macOS the default CJK fonts suffice; on other systems the README directs you to swap the font declarations near the top of book.tex to something like Noto Serif CJK SC. Three releases are published, so compiled artifacts have shipped alongside the sources, and the last push was on 2026-08-10. For readers who only want the content, the released builds are the zero-dependency path; for contributors fixing terminology, the LaTeX source is the working copy.

A terminology policy, written as rules

The translation strategy section is a manual in itself, and it is what keeps a thirty-chapter machine-assisted translation coherent. Universal academic terms stay in English: Transformer, token, attention, agent, policy, reward, embedding, prompt. Personal names, organization names, model names, dataset names and protocol names keep their original forms. Uncommon abbreviations appear as Chinese followed by the English full name and abbreviation, for example supervised fine-tuning rendered with its SVID-style expansion, and terms whose Chinese renderings are not yet common keep the original beside the translation.

Code, citations, references, math environments and TikZ figures are left untranslated, while comments inside code are translated. The README even names the drift class it worries about, terminology inconsistency across chapters, and asks readers to file it as issues, calling it the easiest problem for human review to miss.

The licence chain, and what it obliges

GitHub's licence detector reports the repository as unrecognized, but the situation is documented rather than mysterious: the book is CC BY-SA 4.0, and this translation is a derivative work under the same licence. Practically, the README explains, you may share and adapt it for any purpose including commercial use, on two conditions: attribution to the original author Haggai Roitman and the translator Chasing1020, with the licence link and a note of changes, and share-alike, meaning derived works carry the same CC BY-SA 4.0.

All rights in the original book remain with Roitman. The translator notes the lack of official endorsement twice, in the overview and in the licence section, which along with the AI-assistance disclosure forms a complete provenance statement: origin, transformation, tooling and responsibility are each attributed in a paragraph or less.

Contribution model and limits

The contribution guidance is tailored to a translation project rather than boilerplate. Translation errors and awkward phrasing go to issues or email; improvement pull requests should name the chapter, paragraph and source position; terminology inconsistencies across chapters deserve their own issues; and when the original book updates, readers can claim chapters and sync them. That last point is the project's long-term engine: the translation tracks an upstream that itself evolves.

The limits follow from what it is. It is unofficial, so it can lag the original or carry translation artifacts despite human review; the AI-assistance disclosure means careful readers should treat specific technical renderings with the same skepticism they would apply to any machine-translated text; and the 598-page scope is the original's, not added content. For anyone who prefers the source, the README's citation and licence links lead straight back to Roitman's work.

Against reading the English original and machine-translating it yourself

The obvious alternative is paste-the-chapter machine translation, and this project's existence is an argument against it: a terminology table, consistency review and LaTeX build produce a reference you can cite, while an ad-hoc web translation produces something that drifts mid-chapter. The other alternative is reading Roitman's English original directly, which remains the authoritative text and the right choice for readers comfortable with it; the translation's own rules keep every technical term traceable back to that text anyway.

Against other Chinese-language agentic AI reading lists and handbooks, this volume's claim is completeness and structure: one authored work, six parts, in reading order, with exercises, rather than a curated link farm. For a Chinese-reading engineer who wants the full stack picture in one document, foundations to systems, the translation removes the language tax from a book the field has already taken seriously.

Editorial conclusion

Agentic AI Guide ZH fits Chinese-reading engineers and students who want Roitman's full foundations-to-systems tour in their own language, and contributors willing to fix terminology drift through the project's issue process. Skip it if you need an official, author-reviewed edition, which does not exist, or if you read English comfortably and would rather cite the original. Verify quality on your home turf: download a release or build with XeLaTeX, open the chapter you know best, and judge the terminology against your own field usage before trusting the other 597 pages.

Frequently asked questions

Official sources

1. Chasing1020/agentic-ai-guide-zh on GitHub
2. Issues
3. README
4. Releases