Agent Interview Hub: a Chinese-language interview bank for AI Agent engineer roles

What Agent Interview Hub is, and who it is written for

This is a documentation repository, not a software package. The README describes it as a one-stop interview knowledge base built for engineers preparing to move into AI Agent roles at large Chinese companies, and it names Alibaba, ByteDance, Tencent and Baidu as the target employers. The README also lists 300 or more interview questions, all with answers, requirements and interview write-ups for 14 companies (including OpenAI and Anthropic), six timed practical tasks, and a 16-week study plan.

The audience is narrow and the language is narrower. Every document path in the repository is in Chinese, from 通用知识/Agent核心概念与设计模式.md through the per-company folders such as 阿里巴巴/ and 字节跳动/. If you cannot read Chinese, the questions are of little use, because the value is in the written answers rather than the question titles. The repository is also aimed at interview preparation rather than at building a working agent: the six practical tasks are described as timed challenges modelled on real interview scenarios, and they link back to a separate repository, agent-interview-prep.

How the knowledge base is organised

The README lays out a four-layer taxonomy. The base layer covers Transformer architecture and attention, LLM fundamentals such as tokenization and embeddings, and prompt engineering. The core layer covers agent design patterns (ReAct, Plan-and-Execute, Multi-Agent), the full RAG chain from chunking through reranking to generation, Function Calling and tool use, MCP, mainstream frameworks including LangChain, LangGraph and AutoGen, and context engineering. The advanced layer adds Agentic RAG and GraphRAG, agentic coding tools, fine-tuning and inference optimisation, and agent safety, evaluation and alignment. The practical layer covers system design, company requirements and interview write-ups.

The files themselves are directories of Markdown documents with numbered tables in the README acting as the index. There are 24 numbered entries in the general knowledge section, split into core technology documents, a question bank, and an advanced question set. The company folders are separate trees, each holding requirement pages, question pages and, for some companies, a real interview write-up. The README's table shows dashes for the companies that have no write-up yet, so the coverage is uneven by design rather than by accident.

One structural detail matters for anyone browsing on GitHub: a single canvas file, Agent工程师知识地图.canvas, is meant to be opened in Obsidian. The README states that Obsidian users can open it to review along four lines: the long-term plan, core knowledge, project practice and company interview write-ups. Outside Obsidian, that file is just JSON.

Building the site locally and reading the first document

The README gives two ways to read the material. The recommended route is the hosted GitHub Pages site, which it says is mobile friendly, at zchary1106.github.io/agent-interview-hub/index.html for the classic build and new.html for the exam-preparation layout. The classic page keeps the original content and address and shows a banner pointing to the new version.

For local reading, the README gives a clone and open workflow. There is no server needed to read the Markdown, only an editor.

 git clone https://github.com/Zchary1106/agent-interview-hub.git
 cd agent-interview-hub
 # 用任意 Markdown 编辑器打开即可

The repository does carry a build script for the static site, and requirements.txt pins two packages: Markdown between 3.6 and 4, and pymdown-extensions between 10.12 and 11. The README gives the three commands to install and run it, and states that output lands in dist/.

python3 -m pip install -r requirements.txt
python3 scripts/build_site.py
python3 -m http.server 8000 -d dist

After the third command, the site is served on port 8000 from the dist/ directory. The README also states that GitHub Pages builds and publishes automatically through .github/workflows/pages.yml on every push to the main branch, so a local build is only needed if you want to check rendering before pushing or read offline.

A reasonable first document to open is 通用知识/Agent核心概念与设计模式.md, which the README describes as covering ReAct, Plan-and-Execute and Multi-Agent. From there, 通用知识/八股文完整答案集.md is listed as 69 questions with detailed answers, and 通用知识/RAG核心知识与面试题.md covers the retrieval chain end to end.

The Interview Collector Agent and its install targets

The repository ships an agent definition rather than a running service. The README calls it the Interview Collector Agent and says it searches public sources including Nowcoder, Xiaohongshu, Zhihu, CSDN, Blog Garden, Juejin and GitHub for interview write-ups, turns them into structured candidates, and then syncs them into the index and company documents. The install script takes a target list.

bash agents/interview-collector/install.sh --targets copilot,claude,cursor,generic

The README's table states 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 generic or domestic agents. Note that the Cursor target writes into the current working directory, while the other three write into the home directory.

The collection workflow can also be driven from the command line. The README shows a doctor step that checks local search and reading tools, then a search step that takes repeated platform flags.

# 1. 检查本机搜索/读取工具
python3 scripts/collect_interviews.py doctor

# 2. 搜索公开来源，生成候选 JSON 和 Markdown 报告
python3 scripts/collect_interviews.py search \
  --platform nowcoder \
  --platform zhihu \
  --platform b

The README excerpt cuts off mid-flag at --platform b, so the full platform list is not visible here. Treat the agent as an editorial pipeline that needs a human to review candidates before they reach the index, which is what the phrase structured candidates implies.

Where this repository is the wrong tool

The most obvious limitation is that the README does not name a licence, even though the repository contains a LICENSE file and the README carries a licence badge. Anyone intending to reuse the question bank inside a company, a paid course or a product has no stated terms to point at. That is a gap worth resolving before the content is copied anywhere, and the README does not address it.

Second, this is not an agent framework and cannot be evaluated as one. There is no library to import, no API surface and no runtime. The only executable pieces are a static-site build script and a collection script whose purpose is to gather text. If you came looking for an implementation of ReAct or a RAG pipeline, the documents describe those patterns but do not provide them.

Third, the question bank is a snapshot of a fast-moving field. The README marks several documents as new, including the MCP, Function Calling and Agentic Coding entries, which suggests the maintainers are tracking changes, but the answers themselves carry no version or date stamp in what is visible here. Interview questions about MCP or agentic coding from one hiring cycle can be stale in the next.

Fourth, the company coverage is uneven. The README's own table shows write-ups for Alibaba, ByteDance, Xiaohongshu, Meituan, Ant Group, Huawei, Kuaishou, OpenAI and Anthropic, but dashes for Baidu, Tencent, Google DeepMind, Microsoft and startups. If your target is one of the dashed rows, you get requirements and questions but no first-hand account.

How it differs from a general interview question site

The closest alternative in practice is a general interview question aggregator, and the difference is scope rather than format. A general site covers many engineering roles and treats AI Agent topics as one category among many. Agent Interview Hub inverts that: every document in the tree is about LLM and agent work, and the company folders are organised around AI Agent engineer postings specifically, with requirement pages alongside questions.

The second difference is that the answers are stored in the repository as Markdown rather than behind a login or a paywall, and the README states that the questions all come with answers. That makes the material forkable and reviewable in a pull request, which is also why the collection agent exists: the project treats interview write-ups as something to gather, structure and merge, with an issues link for submitting new ones.

The third difference is the practice layer. The six timed tasks, ranging from a two-to-three hour full-stack exercise to a four-to-six hour multi-agent collaboration task, are closer to a take-home assignment than to a flashcard deck. A question-only site will not give you a four-hour build to attempt under a clock.

Maintenance, upgrade cost and licence status

The repository is not archived, and the last push was on 2026-08-27, which is recent enough that the content is likely still tracking the current hiring cycle. There are no releases, so there is no versioned artefact to pin and no changelog to read before upgrading. Updating means pulling the main branch.

The upgrade cost is low in the technical sense and non-trivial in the editorial sense. The build depends on two Python packages with upper bounds, Markdown below version 4 and pymdown-extensions below version 11, so a future major release of either will not be picked up automatically. Beyond that, there is nothing to migrate: the content is Markdown, and the site is regenerated from it. The real cost of keeping a fork current is re-reading changed answers and deciding whether they still match what interviewers ask.

The licence situation is unresolved. The README includes a licence badge pointing at the LICENSE file, but the licence identifier is not stated anywhere in the repository's visible files, so the terms under which the questions and answers can be reused are unknown. That affects anyone planning to redistribute the content, and it is worth checking the LICENSE file directly before doing so. This is a factual gap, not legal advice.

Editorial conclusion

Use Agent Interview Hub if you are preparing for an AI Agent engineer interview at a Chinese platform company or an overseas AI lab and want questions with written answers plus a weekly plan. Skip it if you need an English-language resource, a runnable library, or a question bank with a stated licence, because the README does not name one. Before relying on it, open the two GitHub Pages builds and compare them, then run the local build with python3 scripts/build_site.py to confirm the dist/ output matches what you expect.

Frequently asked questions

Official sources

1. Issues
2. README
3. Zchary1106/agent-interview-hub on GitHub