# ranxi2001/zero2Agent: 131 articles of agent engineering, and an interview book titled 500 questions that holds 807 > zero2Agent is a Chinese-language agent engineering course that starts from a thirty-line loop and a sixty-line framework and works outward to thirteen modules of framework teardown, training and interview preparation. Its interview book is called Agent Interview 500 Questions while documenting 807 of them across seventeen assessment dimensions. **ranxi2001/zero2Agent** — 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 - Repository: https://github.com/ranxi2001/zero2Agent - Website: https://onefly.top/zero2Agent/ - Stars: 660 · Forks: 43 - Language: Python - License: MIT - Published: 2026-09-10 · Updated: 2026-09-10 · Language: en - Canonical page: https://hysenlabs.com/projects/ranxi2001-zero2agent ## The interview book is titled 500 questions and contains 807 The most-read artefact in this repository has a number in its title that its own description contradicts. The interview book is called Agent Interview 500 Questions, and the PDF file carries that name. The specification beside it says the book covers seventeen assessment dimensions and 807 high-frequency agent and AI interview questions. So the title says five hundred and the contents say eight hundred and seven. Neither number is hidden; they sit within a few lines of each other, which is what makes it checkable. The rest of the book's design is worth reading because it explains how it is meant to be used. The questions are real ones from large Chinese technology companies, with Ant, Alibaba, ByteDance, Tencent, Ctrip and Baidu named. Within a single topic, questions are ordered by how often they appeared in reported interview accounts, which is what makes the book useful on the morning of an interview rather than only in study weeks. Each answer is presented as a beginner answer against an expert answer, so the gap between the two is visible rather than asserted. The book is licensed separately from the repository, under a Creative Commons attribution, non-commercial, share-alike licence with attribution required. ## 131 articles, thirteen modules, twelve finished The progress line in the readme is specific enough to audit, and it holds up. It claims 131 articles across twelve complete modules with one module still being updated. The module table lists thirteen rows with per-module counts, and they sum to exactly 131. The distribution tells you what kind of course this is. The two largest modules are the fundamentals and the interview preparation, seventeen articles each. Next come the framework survey and the harness teardown at thirteen each, then the coding-agent modules at twelve each. The smallest is the SDK comparison at four articles. So roughly a quarter of the material is fundamentals, roughly a quarter is interview preparation, and the middle half is reading other people's agent systems closely enough to explain them. Twelve of the thirteen rows carry a completion marker and one carries an in-progress marker, which is the module on applied work: vibe coding, coding interviews, day-to-day development workflow, automatic harness engineering and evaluation-driven practice. That is the one to expect to change under you, and the reason to read it as a snapshot. ## Two starting points, sixty lines and thirty lines Two of the modules are built by starting from a tiny piece of code and deriving everything else, and they start from different sizes. The framework-derivation module works from about sixty lines of core code. The whole derivation is three lines long, and each line adds one idea to the one above it: ```python workflow = node + node # 有向路径，无循环 chatbot = workflow + loop # 外层循环，多轮对话 agent = chatbot + tools # 图内回路，模型驱动工具 ``` Read as a pipeline: a workflow is nodes wired in a direction with no loop; adding an outer loop makes a multi-turn chatbot; adding tools inside the graph makes an agent whose loop is driven by the model. Three additions, three concepts, and the module then covers retrieval, three different tool forms, memory compression, parallel teams of agents and deployment. The coding-agent module starts smaller still, from a thirty-line agent loop, and works through twelve numbered lessons. The first six cover the loop, tool dispatch, a todo mechanism, subagents, skill loading and context compaction. The second six cover a task dependency graph, background tasks, agent teams, protocols, autonomous agents and worktree isolation. Both modules defer to a separate reference repository rather than reimplementing, which is the honest way to handle other people's code. ## The framework survey counts two things that are not frameworks One module surveys thirteen agent frameworks, and the list is more useful for its inconsistencies than its coverage. Eight entries are recognisable products with a vendor attached: a distributed multi-agent framework from Alibaba, a TypeScript-native workflow framework, Microsoft's plugin-oriented kernel, a Go framework from ByteDance, Google's official agent suite, Microsoft's multi-agent conversation framework with code execution and human in the loop, a Deep Research framework from ByteDance built on a graph library, and a full-stack AI SDK for Next.js applications. Three more are attached to a company but are not frameworks in the same sense: a modular skill system from Anthropic, a harness-engineering entry with no vendor at all, and an agent-universe entry from Huawei. And two have no organisation named whatsoever. One is a code-repository agent pattern with automated pull request review; the other is an entry promising to implement the framework essence from scratch in two hundred lines. That last pair is the most useful thing in the table, because a two-hundred-line reimplementation is a way of teaching what a framework actually does. And the entry for the largest and oldest framework in the space is framed around when not to use it, which is the same instinct. ## A pinned release candidate, and a refusal to treat older material as spec One module is unusually careful about its own epistemic status, and the mechanism is worth copying. The teardown of one agent runtime is framed as design philosophy, treating the system as two things at once: a coding agent you can run directly, and a framework you can reassemble. Its central claim is a negative one, that the system is not a fixed kernel with extension slots, and that the design tries not to keep an irreplaceable privileged core. The official web and headless entry points are described as preset profiles, while the model, tools, filesystem, shell, sandbox, session storage, subagents, interface and even the loop itself can all be swapped through plugins and seams. Then comes the disclosure. The official API is pinned to one specific release candidate version, and the text is explicit that the API serves as evidence for design judgments rather than as the subject. Community tutorials, ebooks, whitepapers and a separate teaching implementation fill in the principles. And the sentence that makes this module trustworthy: where community material is based on an older release, the readme declines to treat it as the current API specification, and points at a third-party notices file that exists in the repository for exactly that purpose. A course that tells you which version it checked, and which versions of other people's writing it rejects, is teaching you something about verification as much as about agents. ## The next module is pinned to a single commit hash The other recent teardown uses a different verification mechanism, and the two together suggest the author treats drift as the main hazard in this field. The coding-agent runtime teardown declares its factual baseline as one specific commit in the official repository, verified on a stated date. Everything in the module is written against that snapshot. The module's stated method is to enter the system through one complete tool round trip rather than through its documentation, and along the way to keep three things apart that documentation tends to blur: runtime events, the persisted transcript, and the model context. The concepts it separates are the ones a framework actually introduces. Provider adaptation is how one interface reaches several models. Skill is described through progressive disclosure, so a capability is loaded only when needed. Extension is a runtime capability, Package is a distribution responsibility, and the model context protocol is assigned a value only in protocol interoperability scenarios, which is a narrower claim than most tutorials make. Every article is said to add reproducible steps, deliberately failing experiments and acceptance evidence, and the module ends by turning a terminal tool back into something recoverable and embeddable through session trees, resume, fork and clone, compaction, remote procedure calls and an SDK. ## Two licences in one repository The repository is Apache-licensed and the interview book is not, and the boundary is stated next to the artefact rather than buried. The book carries a Creative Commons attribution, non-commercial, share-alike licence with a requirement to attribute the source, and the author is credited by a personal profile. The repository itself is MIT. For a course that is largely interview preparation, that split is sensible: the code and prose are meant to be reused freely, while the compiled question bank, which is the commercial and competitive asset, is not. The non-commercial clause on the book is the part to notice. If you were planning to fold the question set into internal training material for a company, that clause is the constraint, and it is the reason the licence sits directly under the download link rather than only in the licence file. There is a third licensing artefact in the repository, a third-party notices file, which covers the community material referenced by the modules rather than the project's own content. ## Twelve module directories, a PDF build, and a custom domain The repository layout says this is a website with source, not a document. Ten of the top-level directories begin with a learning prefix, and the thirteenth module has its own directory rather than a prefixed one. So the module table and the tree correspond exactly, which is not something you can assume. Around them sit the parts of a published site. A domain configuration file pins the site to a custom subdomain rather than the default project pages address. A site configuration file, a layouts directory, a data directory and an index file at the root are the standard pieces of a static site generator. A robots file is present, which is unusual for a course site and suggests the author was thinking about how the material appears in search results. Two more directories earn their place. One holds the PDF build for the interview book, with its output committed alongside it, so the download link points at a file in the repository rather than at a release asset. The other is an examples directory holding a single project, an agent API lab, which is the only executable thing in an otherwise prose repository. The declared primary language is Python, which comes from those scripts and the examples rather than from the articles. ## Conclusion zero2Agent fits a developer who already writes code and uses AI tools but has never taken an agent system to production, and who reads Chinese, since the modules build one architecture at a time rather than surveying many. It is a poor fit if you need English material, since the course is written in Chinese, or if you need to pin a version, because there is nothing here to depend on and the third-party frameworks it covers are moving targets that two modules explicitly pin to a single commit and a single release candidate. Before you start, check three things: which of the thirteen modules answers the question you actually have, since twelve are finished and one is still being written, whether the interview book matches the interviews you expect, given the title and content counts disagree, and how you will handle the framework drift the repository itself flags. ## FAQ ### How many questions are in the zero2Agent interview book? The book's title says 500 questions while its specification says 807 high-frequency agent and AI interview questions across seventeen assessment dimensions. Both numbers appear in the same section of the readme, so the discrepancy is visible rather than hidden. ### How much content is in the zero2Agent course? 131 articles across thirteen modules. Twelve modules are marked complete and one, on applied agent work, is marked in progress. The per-module counts sum to exactly 131, matching the stated progress line. ### What does the zero2Agent framework teardown module start from? About sixty lines of core code, derived in three steps: a workflow of nodes with no loop, adding an outer loop to make a multi-turn chatbot, then adding tools inside the graph so the loop is driven by the model. A separate coding-agent module starts from an even smaller thirty-line agent loop. ### Which agent frameworks does the zero2Agent survey cover? Thirteen entries, including AgentScope, Mastra, Semantic Kernel, Eino, AgentUniverse, DeerFlow, LangChain, Google ADK, the Vercel AI SDK and AutoGen. Three entries have no vendor named: a code-repository agent pattern, a two-hundred-line from-scratch harness implementation, and a harness-engineering entry. ### How does zero2Agent keep its framework teardowns accurate? By pinning them. One module declares a specific release candidate version as its reference and states that the API is evidence for design judgments rather than the subject, declining to treat older community material as the current specification. Another declares a single verified commit hash in the official repository as its factual baseline. ## Sources - [License: MIT](https://github.com/ranxi2001/zero2Agent/blob/main/LICENSE) - [Project website](https://onefly.top/zero2Agent/) - [ranxi2001/zero2Agent on GitHub](https://github.com/ranxi2001/zero2Agent) - [README](https://github.com/ranxi2001/zero2Agent/blob/main/README.md) - [Releases](https://github.com/ranxi2001/zero2Agent/releases) --- Hysen Labs editorial analysis, written from the project's own repository and release notes. Cite the canonical page: https://hysenlabs.com/projects/ranxi2001-zero2agent