ai-agents-from-zero: A Chinese-Language AI Agent Curriculum With Runnable Code

What ai-agents-from-zero Is Trying to Fix

The README states the problem directly: most Chinese-language material on large model application development is fragmented posts or paid bootcamps, and the repository is meant to be a free, systematic alternative. The stated audience is people moving into AI application engineering, agent engineering, or AI automation roles, including front-end, back-end and product people switching tracks. It also targets students preparing for interviews, which is why a separate interview question bank file sits at the repository root next to the chapters.

The scope is wider than a single framework tutorial. The technology table lists LLMs and Transformer internals, prompt engineering and tool calling, the Coze and Dify low-code platforms, LangChain, LangGraph and DeepAgents, the MCP and A2A protocols, RAG with vector stores and reranking, document parsing with MinerU and OCR, deployment with Docker, Ollama, Xinference and vLLM, and fine-tuning with PEFT, LoRA, QLoRA, DeepSpeed and LLaMA-Factory. A repository that claims all of that is either a genuine curriculum or a link farm. The file list suggests the former: 33 numbered Markdown chapters, a glossary, a tools index, and two project directories.

How the Repository Is Laid Out

The reading experience is Docsify. The root holds index.html, _sidebar.md, .nojekyll and sitemap.xml, and package.json defines a single serve script that runs docsify serve . --open. There is no build step and no static site generator; the Markdown files are loaded by the browser at runtime, which is why the online version lives at a GitHub Pages URL. The package.json is marked private and carries only devDependencies: docsify-cli, commitlint, husky and the conventional commit config. Husky runs on prepare, so commit messages are linted, but nothing about the content pipeline is automated.

The Python side is a flat tree of numbered chapter directories rather than an installable package. pyproject.toml declares the project name and a Python range but no build backend and no dependencies; requirements.txt carries the actual pins. That split matters: you cannot pip install this repository, you clone it and install the requirements file. The two project directories, 实战项目-电商问数 and 实战项目-深度研搜, are separate: the README links out to shopkeeper-agent and deepsearch-agents as their own source repositories.

Installing the Examples and Running the First Chapter

The requirements file documents the environment itself in its header comments. The recommended interpreter is 3.10, the supported range is 3.10 to 3.13, and 3.14 is excluded because langchain-redis and similar packages are not yet compatible. pyproject.toml encodes the same ceiling as requires-python = ">=3.10,<3.14".

Create and activate a virtual environment, then install the pinned requirements from the repository root. The commands below are the ones written in the requirements.txt comments.

The Python Version Ceiling Is a Real Constraint

Python 3.14 is excluded, and the reason is stated rather than implied: langchain-redis and other packages in the stack are not yet compatible. If your team standardises on the newest interpreter, you will be pinning down a minor version to use these examples, and that pin propagates into any project you build by copying the code. The upper bound appears in two places, pyproject.toml and the requirements header, which is a good sign that the maintainer treats it as a hard boundary rather than an incidental note.

The second constraint is the flat layout. Chapter directories are examples, not a library. There is no package namespace to import from, no test suite described in the README, and no release artifact. If you want the RAG pipeline from chapter 19 in your own service, you are copying files and adapting them, not adding a dependency. That is normal for a tutorial repository, but it means the maintenance cost lands on you the moment you stop following along and start building.

Where the Curriculum Is Thin and Where It Is Not

The README says the concept chapters are complete and that two projects finished, on 5 May and 17 May. It does not say every chapter is finished, and the numbering runs to 33 with fine-tuning topics at the end. A curriculum that keeps adding chapters as the ecosystem moves will always have a trailing edge, so the practical move is to open the chapter you need and check whether it contains prose or only a title.

The interview question bank is the part that is hardest to evaluate from the outside. The README describes it as organised by job capability area with question and answer formats, cross-linked to the main text by question number, and says a substantial portion comes from real interview questions, public interview write-ups and follow-up scenarios at large companies. That is a claim about provenance, not a guarantee of currency, and interview questions age faster than framework APIs. Treat it as a study aid for structuring your own answers rather than a source of truth.

Licensing is MIT, which is permissive for the text and the example code. The README carries a sponsorship block for a cloud GPU provider with a referral link and a five yuan credit. That is an advertisement inside the documentation, and you should read it as one.

Comparing It With the Official Framework Docs

The obvious alternative is the upstream documentation for LangChain and LangGraph themselves. The difference in approach is real: the framework docs are reference material organised by API surface, they are in English, and they are versioned against the library you actually install. This repository is organised by learning sequence, in Chinese, and it deliberately spans competitors that no single vendor would document together, pairing Coze and Dify with LangChain and LangGraph and adding a fine-tuning track that the agent frameworks do not cover at all.

That breadth is the reason to pick it and also the reason it cannot replace the reference docs. When a LangGraph API changes, the upstream docs change with the release; this repository changes when someone rewrites the chapter. The README says the outline is not fixed and that new knowledge points and frameworks will be merged into the route, which is an honest description of a moving target. Use it to decide what to learn and in what order, then verify the API details against the version you installed.

Who Should Adopt It and What to Check First

The repository is not archived, and the last push was on 2026-06-23, so the content was being written within the last three months. There are no releases and no version tags, so upgrade cost is not a concept that applies here: you pull the branch and the chapters change underneath you. If you have copied example code into a project, pin your own dependencies and treat the repository as a one-time source rather than something you track.

Adopt it if you read Chinese, want a single ordered path from prompting through RAG, LangGraph, MCP and LoRA fine-tuning, and prefer running examples over reading API reference. Skip it if you need English material, if you want a library you can depend on, or if your environment is pinned to Python 3.14. Before you commit a weekend, open the chapter file for the topic you care about and check that it has body text, and read .env-example to confirm which provider keys the examples expect.

Editorial conclusion

Adopt it if you read Chinese and want one repository that spans prompting, RAG, LangGraph, MCP and LoRA fine-tuning with example code you can run locally on Python 3.10 to 3.13. Skip it if you need English material or a maintained library rather than a course, since it is a teaching repository with no releases and no versioned API. Before committing time, check whether the specific chapter you need is finished: the README says the concept chapters are complete and names two finished projects, but it does not claim every chapter is done, so open the chapter file and confirm it has body text rather than a heading.

Frequently asked questions

Official sources

1. didilili/ai-agents-from-zero on GitHub
2. Issues
3. License: MIT
4. Project website
5. README