didilili/ai-agents-from-zero: Chinese MIT AI Agent Guide

June 13, 2026·AI Newsroom·8 min read·9 sources
open-sourcelangchainlanggraphmcpragskills+5
ai-agents-from-zero open-source repository social preview card
Image: GitHub / didilili/ai-agents-from-zero repository (MIT)

On June 13, 2026, the public repository didilili/ai-agents-from-zero — published by the Chinese Datawhale organization under the MIT license (LICENSE, copyright 2026 didilili) — counts 1,914 stars, 254 forks, 27 markdown chapters, two completed projects released in May 2026, and an AI interview question archive. The README, in Chinese, presents itself as “the most systematic AI agent crash course on the web (from zero to enterprise deployment)” (README, June 13, 2026). All chapter code and both projects are in Python, making them readable independent of the README language: the language barrier applies to prose, not samples.

What happened

The repository covers, in a single MIT syllabus, the full AI agent stack that readers are adopting in 2026: LLM fundamentals (chapters 1, 8, 11), low-code with Coze and Dify (chapters 3-6, with dedicated chapters on Python API calls and Windows/Linux deployment), LangChain v1.x (chapters 9-19: Model I/O, Prompt, Parser, LCEL, Memory, Tools, RAG, Agent), MCP (Model Context Protocol) in a dedicated chapter (chapter 20) comparing with Function Calling and custom servers, LangGraph (chapters 22-26: graphs, State, Node, Edge, persistence, multi-agent, A2A protocol), Skills and AI programming tools (chapter 27, covering Cursor / Codex / Claude Code / Trae / Qoder), advanced RAG (vector DB, sparse + dense retrieval, BGE-Rerank, HyDE, RAGAS evaluation, MultiAgent for e-commerce NL2SQL), and fine-tuning (PEFT, LoRA, QLoRA, DeepSpeed, Llama-Factory, Alpaca / ShareGPT formats). The chapter 9 title is 9-LangChain概述与架构.md, chapter 22 is 22-LangGraph概述与快速入门.md, and the interview question bank is AI智能体与大模型应用开发面试题库.md (134,617 bytes, organized by competency domain) (repo contents, June 13, 2026).

Repository numbers from the GitHub REST API on June 13, 2026: 1,914 stars (1.9k rounded), 254 forks, 10 open issues, MIT license (SPDX), primary language Python, size 681.8 MB, last metadata update 2026-06-13 03:34:53 UTC, last push 2026-06-10 02:55:31 UTC, public repository created 2026-01-29. Declared topics: agent, agent-framework, agentic-ai, ai-agent, aigc, coze, cursor, deepagents, dify, gpt, langchain, langgraph, llm, mcp, rag, skills, tutorial.

The two completed projects are released as separate sub-repositories and linked from the main README, with explicit completion dates in the changelog (教程更新日志.md):

The documentation site is published at didilili.github.io/ai-agents-from-zero/ (HTTP 200, 22,046 bytes, language zh-CN); it’s a Docsify instance with sidebar navigation, learning map, and rendered chapters. The README discloses sponsorship from UCloud AI cloud for the “Agent Plan” package — honest and visible disclosure (“感谢 优云智算 赞助本项目”), not a masked operation.

Why it matters

1. The stack coverage aligns with the 2026 production AI agent stack. A single guide touching MCP + LangGraph + RAG + Skills + fine-tuning under MIT, in one syllabus, is rare. Open resources on these topics are usually scattered across 5-10 separate repositories (e.g., LangChain Python docs, LangGraph docs, MCP spec, separate RAGAS tutorials, separate Llama-Factory tutorials). Having chapters 1-27 in one place, with the same style and a single glossary system (全书术语表.md, 62 KB), lowers orientation costs for anyone building a multi-component architecture.

2. The two completed projects are the real value, not the theory. 电商问数 (NL2SQL on MySQL+LangGraph+Qdrant+Elasticsearch) and 深度研搜 (DeepAgents multi-agent with WebSocket live progress) are runnable: copy them, run docker compose locally, read the Python code, and you get a real RAG + multi-agent + REST + WebSocket pipeline. For a team wanting to see “how you put this kind of system together in production,” they’re a rare open-source case study, and the completion dates (May 3 and May 17, 2026) make them current.

3. The language barrier should be explicitly acknowledged. The README, sidebar, 27 .md files, and interview question bank are in Chinese (zh-CN, with Chinese filenames: 9-LangChain概述与架构.md, 20-MCP模型上下文协议.md, AI智能体与大模型应用开发面试题库.md). The code samples are in Python, readable by anyone who knows the language. An article promising “it’s in English” would be wrong as of June 13, 2026: the main branch is Chinese. An article saying “the code reads, the prose doesn’t” is honest and lets readers decide if it’s worth their translation time (or if reading just the Python source is sufficient).

4. The “long tutorial + interview question archive + real projects” pattern is a Chinese industrial format. The interview question bank aligns with the “AI Agent / LLM Application Engineer” job description profile, a standard professional figure in China in 2026. For an Italian candidate preparing for AI engineer interviews, it’s a complement to the classic “Cracking the FAANG Interview”: it covers areas (MCP, A2A, multi-agent, eval & observability) that anglophone collections treat less. Not a substitute — a specialist supplement.

5. It’s a supplement, not a substitute for official documentation. LangChain, LangGraph, MCP have their own well-maintained docs, in English, often more up-to-date on current API details. The value of didilili isn’t “the definitive guide” — it’s “a long, organized, Chinese guide that covers the full stack and lets Python readers use it as a cross-reference even without translating the prose.” The article presents it for what it is.

What to watch

Risks and caveats

  1. The README is in Chinese, not English. Verified June 13, 2026 on the main branch via REST API and raw README rendering. An official or unofficial English translation is not yet part of the canonical branch.
  2. The syllabus aligns with the Chinese 2026 “AI Agent / LLM Application Engineer” job profile. It’s not automatically transferable to the Italian or EU market. Job titles, interview expectations, and thematic cuts reflect that market, not the European one.
  3. “Complete” means “27 chapters + 2 projects,” not “comprehensive.” It’s not the definitive guide to AI agents: the README itself calls it a “速成指南” (crash course). It’s a systematic starting point, not an exhaustive reference.
  4. The MIT license permits forking and redistribution with attribution, but the article doesn’t cite third-party forks unless they’re verifiable primary sources. The deepsearch-agents sub-repo doesn’t have a visible LICENSE file in GitHub metadata as of June 13, 2026 (while shopkeeper-agent has MIT); for the sub-repo, the principle of not making unverified license claims applies.
  5. Star/fork/commit numbers grow over time. 1,914 stars and 254 forks were observed on June 13, 2026; this isn’t a permanent snapshot. Observation dates should be cited for every number.
  6. It’s not an installable library. The repository has no PyPI or npm releases. It’s a content + code repository to clone or read, not pip install.
  7. The repository has declared sponsorship. UCloud AI cloud (优云智算) is explicitly cited in the README as a sponsor. It’s an honest disclosure, not editorial bias, and the article flags it as transparency — not an endorsement of the service.
  8. The article doesn’t compare the guide to official LangChain tutorials, LangGraph tutorials, DeepAgents tutorials, MCP spec, RAGAS, or paid courses in terms of “quality” or “completeness.” Those comparisons aren’t factual without measurable metrics; the article is limited to what’s verifiable (“covers MCP, LangGraph, RAG, and Skills in a single MIT syllabus”).

What to do

For AI engineers working with LangChain / LangGraph / MCP / RAG / Skills: the repository is a cross-reference to be used alongside — not to replace — official documentation. The documentation site sidebar and learning map are recommended entry points. Python code samples are readable without Chinese: open them, copy them, run them locally for lowest friction. The two completed projects (shopkeeper-agent for NL2SQL and deepsearch-agents for DeepAgents multi-agent) are runnable: clone them, read the docker compose, understand their structure. Cost: time, not license.

For those preparing for AI engineer technical interviews: the interview question bank is a complement to “Cracking the Coding Interview” / LeetCode, with specialist coverage of MCP, A2A, multi-agent, evaluation, and observability. Use it as a supplementary question bank for areas less covered by anglophone collections. Not a substitute — a targeted extension. Here too: the Python sample code reads, the theory questions need translation.

For those looking for a Chinese open-source alternative to anglophone guides: the “systematic long tutorial + JD-aligned interview archive + complete real projects” pattern is a didactic format Datawhale has industrialized. It’s worth flagging as an editorial signal: the Chinese open-source AI ecosystem is, on certain fronts (full-stack, JD alignment, MIT-covered completeness from a single syllabus), more methodical than anglophone sources suggest. Not a qualitative verdict — a format observation. For an Italian publisher or educator, it’s an opportunity to broaden the teaching material selection.

Verdict

didilili/ai-agents-from-zero is a high-quality, well-organized open-source guide, published under the MIT license, with real code examples, two completed projects, and an AI interview question archive structured by competency domain. It’s not the definitive guide to AI agents, and doesn’t claim to be. It’s one of the most systematic open resources on MCP + LangGraph + RAG + Skills in 2026, with end-to-end coverage that’s hard to find in a single repository.

The value for readers depends on tolerance for the language barrier. Python readers can use the repository as a cross-reference even without translating the prose; those wanting only the prose will find Chinese the first wall, and will need to decide if it’s worth the translation time (for the sidebar and learning map, automatic translation is sufficient; for denser chapters, translation may require multiple passes). In both cases, the repository deserves to be known: it’s not a substitute for official LangChain, LangGraph, MCP documentation — it’s a serious, MIT, Chinese, code-readable addition. The Chinese open-source AI ecosystem deserves to be read even just as a signal of how agent training is being industrialized in 2026.