agent-interview-hub: A Chinese-Language Interview Bank for AI Agent Engineering Roles
AI Agent 工程师面试资料库 - 国内大厂面经、岗位要求、高频面试题
At a glance
- What is it?
- A Markdown knowledge base plus static site generator, aimed at engineers preparing for Agent roles at Chinese internet companies and, in a newer section, at OpenAI, Anthropic, Google and Microsoft. The content is the product; the tooling around it is thin, and the repository states no licence.
- Who is it for?
- Adopt this if you are interviewing for an Agent or LLM application role where the loop is conducted in Chinese and you need answer-backed question sets rather than a tutorial. Do not adopt it as a production dependency or as your only English-language source, and do not redistribute the content until the repository states a licence.
- 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 HTML, 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 this repository fills: answer-backed questions for Chinese Agent interviews
Most LLM interview material available in English is either a course or a blog series, and neither is organised around what a hiring loop actually asks. This repository inverts that. Its README describes 300+ questions with answers attached, job requirements and interview write-ups for 14 companies, six timed practical exercises, and a 16-week study plan. The stated audience is engineers preparing to change jobs into Agent positions at Alibaba, ByteDance, Tencent and Baidu, with a newer section covering OpenAI, Anthropic, Google DeepMind, Microsoft and startups. The knowledge taxonomy in the README runs from Transformer attention and tokenization up through ReAct, Plan-and-Execute and Multi-Agent patterns, RAG chunking and reranking, Function Calling, MCP, LangChain and LangGraph, then to evaluation, safety and inference optimisation. That ordering is the useful part: it tells you what the maintainer thinks a loop covers, and it is broad enough that a reader can locate their own weak area without reading the whole tree. The repository is a study corpus, not a library. Nothing here runs at inference time, and nothing here is imported into your code.
Repository layout: Markdown files, company folders, and two HTML entry points
The primary language is HTML, which is slightly misleading. The content lives in Markdown: a top-level file named Agent工程师学习路线图.md for the 16-week plan, a 通用知识 directory holding the numbered technical documents and question banks, a 项目实战/实操考题 directory holding the six timed exercises, and one folder per company (阿里巴巴, 字节跳动, 小红书, 百度, 腾讯, 美团, 蚂蚁集团, 华为, 快手, OpenAI, Anthropic, 谷歌, 微软, 初创公司). Each company folder is documented as containing up to three artefacts: 岗位要求, 面试题 and 真实面经. The README's own table shows this is uneven. Alibaba, ByteDance, Xiaohongshu, Meituan, Ant Group, Huawei, Kuaishou, OpenAI and Anthropic have all three; Baidu, Tencent, Google, Microsoft and the startup folder have no 真实面经. Treat that column as the honest signal of coverage depth. Two HTML pages are published: index.html, described as the classic version, and new.html, described as a job-preparation layout with a banner on the classic page pointing readers to the new one. There is also an Obsidian canvas file, Agent工程师知识地图.canvas, for readers who want to trace four lines (long-term route, core knowledge, project practice, company write-ups) in that tool.
Getting it running locally: clone, build, serve
Reading the material requires nothing beyond a Markdown editor. The README gives the clone sequence as git clone https://github.com/Zchary1106/agent-interview-hub.git followed by cd agent-interview-hub. To reproduce the published site, the documented steps are python3 -m pip install -r requirements.txt, then python3 scripts/build_site.py, then python3 -m http.server 8000 -d dist. Output lands in dist/. Publication is handled by .github/workflows/pages.yml, which the README says builds and publishes on push to main. Two caveats worth stating plainly. First, the build depends on requirements.txt, whose contents are not shown in the material available here, so the Python version and dependency set cannot be confirmed. Second, the repository has no releases and no homepage field, so the only supported path is building from main. If you want the site without building it, the two GitHub Pages URLs in the README are the shorter route.
The Interview Collector Agent and its install targets
The one piece of executable tooling described in the README is an Interview Collector Agent that searches public sources (牛客, 小红书, 知乎, CSDN, 博客园, 掘金, GitHub) for interview write-ups and turns them into structured candidates, which are then synced into the index and company documents. It installs with bash agents/interview-collector/install.sh --targets copilot,claude,cursor,generic, and the README lists exactly where each target writes: ~/.copilot/instructions/interview-collector.instructions.md for GitHub Copilot CLI, ~/.claude/skills/interview-collector/SKILL.md for Claude Code, .cursor/rules/interview-collector.mdc for Cursor, and ~/.agent-interview-hub/interview-collector/AGENT.md for a generic or domestic agent. A scripted path also exists: python3 scripts/collect_interviews.py doctor checks local search and read tools, and python3 scripts/collect_interviews.py search takes repeated --platform flags such as nowcoder and zhihu to produce candidate JSON and a Markdown report. Note the design decision here: the collector writes into your home directory or your project's .cursor directory, not into a sandbox. It is a contributor tool for maintaining the corpus, not something a job candidate needs. Most readers will never run it.
No licence, no releases, and a corpus that ages quickly
The licence is unknown. The README carries a LICENSE badge, but the repository metadata does not identify a licence, and the material available here does not show the file's contents. That matters more than usual for this kind of project. The content is described as sourced from public posts on 牛客, 小红书, 知乎, CSDN and elsewhere, and interview write-ups are personal accounts. Without a stated licence, the default position is that all rights are reserved, so copying the question banks into an internal wiki or a paid course is not something the repository grants you permission to do. This is a description of the repository's stated position, not legal advice; if redistribution matters to you, read the LICENSE file and decide for yourself. The second limitation is freshness. Interview questions track what companies ask, and what companies ask tracks the framework churn visible in the README itself, where Function Calling, MCP, Agentic Coding and Agent Harness documents are all marked as new additions. A bank that was current at one push can lag within a couple of hiring cycles. The repository has no releases, so there is no version to pin and no changelog to read; the last push timestamp is the only signal of activity you get.
How it differs from DataWhale and from framework documentation
The README points to a DataWhale community question bank as one of its own sources, which makes the comparison concrete rather than hypothetical. DataWhale is a community-maintained open-source study organisation whose material is oriented toward learning a subject from the ground up. This repository is oriented toward passing a specific loop: it pairs questions with answers, attaches company job requirements, and adds a per-company 真实面经 column. The difference in approach is scope versus targeting. If you want to understand how RAG reranking works, the framework documentation and DataWhale material will serve you better, because they are written to teach. If you want to know which questions a ByteDance loop is reported to ask and what a plausible answer looks like, this repository is organised for that and the other resources are not. The trade-off is that a targeted bank can only be as good as its sourcing, and the README does not state how individual write-ups are verified. Cross-check any company-specific claim against a second source before you build an interview story around it.
Who should adopt it, and what to check before you rely on it
This fits an engineer with an LLM application background who is preparing for a loop conducted in Chinese and wants a structured question set with answers, a company-by-company view of requirements, and a weekly plan to work through. It also fits a team lead who wants a shared reading list for engineers moving into Agent work, provided the licence question is resolved first. It does not fit someone who needs an English-language resource, someone preparing for a research or model-training role rather than an application engineering role, or anyone who wants a maintained dependency: there are no releases, no package on any registry, and the build path assumes you can satisfy requirements.txt. Before committing study time, check three things. Open the LICENSE file and confirm what the repository actually permits. Run the documented build (python3 -m pip install -r requirements.txt, then python3 scripts/build_site.py) and confirm it succeeds on your Python version, since the requirements are not visible in the material reviewed here. Then open the company folder you care about and count the files: if 真实面经 is absent, as the README's own table shows for Baidu, Tencent, Google, Microsoft and the startup folder, you are reading a requirements summary and a question list, not a first-hand account of the loop.
Editorial conclusion
Adopt this if you are interviewing for an Agent or LLM application role where the loop is conducted in Chinese and you need answer-backed question sets rather than a tutorial. Do not adopt it as a production dependency or as your only English-language source, and do not redistribute the content until the repository states a licence. Verify first: open LICENSE, confirm the build path with python3 -m pip install -r requirements.txt followed by python3 scripts/build_site.py, and check whether the company folder you care about has a 真实面经 file or only 岗位要求 and 面试题.
Community notes