zero2Agent: A Chinese-Language Agent Engineering Curriculum With a 630-Question Interview Greenbook
面向大厂Agent研发岗位求职的agent教程网站，涵盖技术路线与面试八股文
At a glance
- What is it?
- zero2Agent is an MIT-licensed tutorial repository and static site that teaches Agent engineering from model APIs up to a hand-written harness, and ships a separate CC BY-NC-SA interview PDF. The Python code is scaffolding for the prose, not a library you install.
- Who is it for?
- Adopt zero2Agent if you already write code and want a structured path from Tool Calling through StateGraph to a hand-written harness, and if reading Simplified Chinese is not a barrier. Do not adopt it if you need an installable Agent runtime, an English-language course, or code you can drop into production: nothing in the README describes a published package.
- Can I use it commercially?
- Yes. MIT is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
- Is it still maintained?
- Yes. The repository last received commits 1 day ago.
- What is it written in?
- Mainly Python, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on September 15, 2026, and from our analysis. They are not legal advice.
DEEP OPEN-SOURCE ANALYSIS
The Gap zero2Agent Targets: Developers Who Can Call an API but Cannot Ship an Agent
The README states the audience directly: programmers who already write code and use AI tools but have not systematically built an Agent system, plus new graduates and career switchers preparing for Agent roles. That is a narrower group than "people who want to learn AI." The stated goal is to move past demos, prompts and wrapper workflows into engineering design, framework teardown and code implementation. The repository is a curriculum, not a toolkit. Its primary language is Python, but the material spans TypeScript and Go frameworks, and the deliverable is a reading path plus a PDF. The 131 articles are organized into 12 modules, one of which (Agent 应用实战) is still marked in progress. If you want a library to import, this is the wrong repository. If you want a sequence of lessons that ends with a deployable project, the README claims exactly that: the Final Project module builds an interview-diagnosis Agent called OfferPilot through a ten-layer harness, from PRD to deployment.
How the Curriculum Is Layered: Concepts, Then Graphs, Then Hand-Written Harnesses
The module table reveals a deliberate progression. Agent Basic (17 articles) covers model API input and output, message traces, Tool Calling, Memory patterns, context and KV/Prompt Cache, compression, and why a demo runs while a production deployment does not. LangGraph (7 articles) then introduces StateGraph through TypedDict state design, node functions, compilation, add_edge sequencing, add_conditional_edges routing, parallel fan-out and fan-in, prompt chaining and LLM integration. From there the repository branches into framework surveys and source-level teardowns: OpenClaw derives an Agent from a 60-line core using the progression node plus node equals a workflow, workflow plus loop equals a chatbot, chatbot plus tools equals an agent. Claude Code spends 12 lessons rebuilding a coding agent from a 30-line loop through tool dispatch, TodoWrite, subagents, skill loading, context compaction, task DAGs, background tasks, agent teams, protocols and worktree isolation. The DeepSeek Harness module is the most opinionated: it describes a runtime with no privileged, non-replaceable core, where the model, tools, filesystem, shell, sandbox, session store, subagent, UI and even the loop are replaceable through plugins and seams. That is a design argument, and the README is explicit that the official API is pinned to dsh-v0.1.0-rc.8 while some community material is based on the earlier rc.6.
Getting the Material: A Static Site and a PDF, Not a pip install
There is no installation step described in the README. The entry points are the GitHub Pages site at onefly.top/zero2Agent and per-module URLs such as /learn-agent-basic/, /learn-langgraph/, /learn-sdk-frameworks/, /learn-agent-survey/, /learn-openclaw/, /learn-claude-code/, /learn-deepseek-harness/, /learn-pi/, /learn-codex/, /learn-agent-training/, /learn-agent-practice/, /learn-agent-interview/ and /final-project/. The interview material is distributed as a PDF committed to the repository at publish-pdf/output/zero2Agent-绿皮书-Agent面试500问.pdf. The README's release notes list v1.0.0 (2026-08-22), v1.1.0 (2026-08-25) and v1.2.0 (2026-08-26) for that greenbook, with the last push to main on 2026-09-10. If you want to work through the code examples locally, you clone the repository and follow the article text; the README does not document a requirements file, a virtualenv step, or a test command, so treat any local execution as something you configure yourself.
The Interview Greenbook Is a Separate Artifact With a Separate Licence
This is the detail most readers will miss. The repository is MIT, but the greenbook PDF carries its own terms: the README states CC BY-NC-SA 4.0 and asks for attribution. Non-commercial and share-alike are materially different from MIT. If you plan to redistribute the PDF inside a company, or bundle it into a paid course, the licence on that file is the one that applies, not the repository LICENSE. The content itself is described as 17 assessment dimensions and 630 Agent and AI interview questions drawn from named employers, with each topic ordered by how often it appears in real interview accounts, and each answer presented as a novice answer versus an expert answer. That format is useful for calibration, and it is also the format most likely to go stale: interview questions track hiring fashion, and a PDF pinned to a release date does not update itself. The README gives no cadence for greenbook revisions beyond the three releases in August 2026.
Where the Curriculum Is Thin, and Where It Will Age Fastest
Two modules carry visible risk. Agent 应用实战 is marked in progress, so the practice material on Vibe Coding, AI Coding interviews and Eval-Driven workflows is incomplete by the repository's own accounting. The framework survey module is broader but shallower by construction: 13 articles covering AgentScope, Mastra, Semantic Kernel, Eino, GitAgent, Harness engineering, AgentUniverse, DeerFlow, LangChain, Google ADK, Skills plus Claude Code, Vercel AI SDK and AutoGen. Thirteen frameworks in thirteen articles means each gets roughly one article's worth of depth, which is enough to decide what to read next and not enough to debug a production issue. The DeepSeek Harness module is the sharpest example of version drift: the README pins the official API to dsh-v0.1.0-rc.8 and explicitly declines to treat community material based on rc.6 as the current specification. A release candidate as the reference point means the module can be correct on the day it was written and wrong a month later. A third limitation is linguistic. Despite the English repository name and description, the README, module names and article content are in Simplified Chinese. There is no indication of an English translation.
Compared With LangChain's Documentation and DeepLearning.AI-Style Courses
The nearest alternative for a working developer is the official documentation of whichever framework you intend to use, plus a structured course. LangChain's own docs, for instance, are written by the maintainers, versioned alongside the code, and updated with each release; they are also scoped to one framework and assume you have already chosen it. zero2Agent inverts that: it surveys 13 frameworks before committing, and its LangGraph module teaches the graph primitives rather than the framework's full API surface. The trade-off is maintenance. LangChain's docs track LangChain. A 131-article independent curriculum tracks every framework it mentions, and the README's own handling of the DeepSeek Harness rc.6 versus rc.8 discrepancy shows how much editorial work that requires. The second alternative is a paid video course, which offers a paced schedule and a cohort. zero2Agent offers neither; it is text, self-paced, and free apart from the licence constraints on the PDF. If you learn better from lectures or need someone to grade your work, this repository will not replace that.
Maintenance Cost and What to Verify Before You Commit
The repository is not archived, and the last push recorded is 2026-09-10, roughly two weeks after the v1.2.0 greenbook release. Three greenbook releases in four days in August 2026 suggest an active editorial push rather than a steady cadence, and the README does not promise a schedule. For a reader, the maintenance cost is not in your dependency tree, it is in re-reading: any module pinned to a release candidate or a specific SDK version will need checking against upstream before you rely on it. The concrete items to verify first are the DeepSeek Harness API pin (dsh-v0.1.0-rc.8) against the current upstream release, the status of the Agent 应用实战 module, and whether the greenbook PDF's CC BY-NC-SA 4.0 terms fit how you intend to use that file. The repository's THIRD_PARTY_NOTICES.md is cited in the README as the place where source and licence provenance for third-party material is recorded; read it before reusing any of the surveyed content.
Editorial conclusion
Adopt zero2Agent if you already write code and want a structured path from Tool Calling through StateGraph to a hand-written harness, and if reading Simplified Chinese is not a barrier. Do not adopt it if you need an installable Agent runtime, an English-language course, or code you can drop into production: nothing in the README describes a published package. Before committing, open the Agent Basic module and one harness module, and confirm that the pinned API version in the DeepSeek Harness module still matches the upstream release you intend to use.
Community notes