AgentGuide: A Structured Path to AI Agent Engineering and Interview Readiness

What AgentGuide Solves and Who It Is For

Engineers entering the AI Agent field often find themselves with scattered resources: isolated tutorials on LangGraph, separate blog posts on RAG, disconnected interview prep for large model positions. AgentGuide addresses this by organizing the full preparation path, from understanding what an Agent Loop is to completing interview-ready projects, into a single repository.

The target reader is a software engineer or researcher in China who wants to move into AI Agent development or algorithm engineering at a company that works with large language models. The repository explicitly positions itself as doing for AI Agent job seekers what JavaGuide does for Java developers: providing a system that is both technically comprehensive and oriented toward career outcomes, not just academic learning.

Two Tracks: Algorithm Engineer and Development Engineer

The repository offers two distinct learning paths that share a common foundation but diverge in emphasis.

The algorithm engineer track emphasizes theory, experimentation, and training methods. It covers Agent reasoning and planning (ReAct, Reflexion, Tree and Graph Search), tool-use strategies, RAG algorithms including Hybrid Retrieval, Rerank, GraphRAG, and Agentic RAG, and post-training methods including SFT, LoRA, QLoRA, DPO, and GRPO. The track culminates in understanding trajectory data synthesis and Reward or Verifier design for Agentic RL.

The development engineer track focuses on building systems that run reliably in production. It covers the Agent Harness: state management, tool registration, permission confirmation, sandbox, trace, replay, and cost control. It also covers tool protocols (MCP, Skills, A2A, ACP, API adapters, Browser and Computer-use tool wrapping) and production-grade RAG with document parsing, vector databases, reranking, citation, observability, CI evaluation, and safety red-teaming.

The README decision tree for choosing between the two tracks is the correct starting point. The same project can be framed as an algorithm submission or an engineering submission depending on which track the reader is following.

Repository Layout and How to Navigate It

The top-level directory contains docs/, examples/, apps/, projects/, and resources/. The docs/ folder is organized by topic area, with subdirectories for getting-started guides, roadmaps, research frontiers, and interview preparation. The examples/ directory contains files that illustrate the repository conventions.

examples/eval-cases.jsonl holds a sample evaluation case set showing the format for structuring test inputs and expected outputs. examples/minimal-agent-loop.md is a walkthrough of the minimum components needed for an Agent Loop. examples/project-readme-template.md is a template for documenting projects in a way that makes the architecture and trade-offs clear for resumes and interviews. examples/tool-card-template.md gives a format for documenting tool definitions. examples/trace-schema.json provides a JSON schema for Agent execution traces.

The README lists eight learning outcomes that the repository structures its content around: architectural understanding (distinguishing Chatbot, Workflow, Agent, and Multi-Agent), engineering capability (LangGraph, OpenAI Agents SDK, Context Engineering, Memory, Tools, MCP, and Skills), RAG design (Hybrid Retrieval, Rerank, GraphRAG, Agentic RAG, Multimodal RAG), reliability engineering (Eval Set, Trace/Replay, LLM-as-a-Judge, Sandbox, HITL, permission and cost control), training methods (SFT, DPO/GRPO, trajectory data synthesis, Reward and Verifier), project delivery (runnable, evaluable, reproducible projects with architecture documentation), interview communication (answering follow-up questions on principles, system design, and engineering trade-offs), and career path clarity (algorithm vs development role differences). This checklist defines what the repository is designed to produce, not what it contains on its own.

The README lists these eight outcomes in a code block:

✅ 【架构认知】分清 Chatbot、Workflow、Agent 与 Multi-Agent，理解 Agent Loop 和 Harness
✅ 【工程能力】掌握 LangGraph / OpenAI Agents SDK、Context Engineering、Memory、Tools、MCP 与 Skills
✅ 【RAG 能力】能设计 Hybrid Retrieval、Rerank、引用溯源、GraphRAG、Agentic RAG 与 Multimodal RAG
✅ 【可靠性】会构建 Eval Set、Trace / Replay、LLM-as-a-Judge、Sandbox、HITL、权限和成本控制
✅ 【训练认知】理解 SFT、DPO / GRPO、工具调用与轨迹数据合成、Reward / Verifier 的基本方法
✅ 【项目交付】完成 2-3 个可运行、可评测、可复现的项目，并写清架构、指标、取舍与失败分析
✅ 【面试表达】能围绕原理、系统设计、实验结果和工程权衡回答追问，而不是背“标准答案”
✅ 【求职路径】明确算法岗与开发岗能力差异，用项目、开源贡献和技术内容构建作品集

Core Technical Coverage Across the Full Stack

The README documents coverage across three layers: Agent application, Agent Harness, and Data/Eval/Training.

At the application layer, the repository covers frameworks including LangGraph, OpenAI Agents SDK, AutoGen, CrewAI, Pydantic AI, Dify, n8n, and Flowise. Task types include Research Agents, Coding Agents, Web Agents, Multi-Agent systems with Supervisor and Handoff patterns, and Computer Use with Browser Automation.

At the Harness layer, the focus is on Context Engineering: System, Memory, Retrieval, and Trace context, along with Context Compression, Prompt Cache, and cost optimization. Tool protocols covered include Tool Schema, MCP, Skills, A2A, ACP, and permission tiering. Reliability topics include Sandbox, HITL (Human in the Loop), Retry, Cost Guard, Trace, Replay, and Observability.

At the Data and Eval layer, the RAG section covers document parsing tools (Docling, MinerU, Unstructured), vector databases (Milvus, Qdrant, Chroma, FAISS), and retrieval strategies including GraphRAG, Agentic RAG, and Multimodal RAG. Evaluation frameworks mentioned include Promptfoo, DeepEval, Inspect, and RAGAS. Benchmark environments include WebArena, OSWorld, and SWE-bench. Post-training methods span SFT, LoRA, QLoRA, DPO, GRPO, and tool-use or trajectory data synthesis.

The breadth is notable. An engineer following either track is expected to understand the full stack, not just the layer their role focuses on.

The Job-Seeking Framework and Portfolio Requirement

The README frames job seeking as a distinct skill from technical learning and dedicates a section to what it calls a new paradigm: doing matters more than studying. The concrete output expected at the end of the preparation path is two to three projects that are runnable, evaluable, and reproducible, with written documentation of the architecture, metrics, trade-offs, and failure analysis.

The repository includes a companion project called learn-workbuddy, which the README describes as a clean-room teaching reproduction of a WorkBuddy-style Desktop Agent Harness. It covers Agent Loop, tool calling, Context Engineering, long-term memory, Sidecar architecture, permission auditing, and model evaluation. This is the closest AgentGuide comes to providing executable starting code; it is a separate repository, not code bundled in AgentGuide itself.

The README claims an 8-10 week timeline from start to interview readiness, and a 2-3 week timeline to complete a resume-quality project. These represent the author stated targets, not measured outcomes across a cohort.

Limitations and What the Repository Does Not Provide

AgentGuide aggregates resources and provides a structured path, but it does not replace the underlying materials. Much of the learning it prescribes requires reading papers, external courses, and third-party tutorials that the repository links to but does not reproduce. Engineers expecting a self-contained textbook will need to follow those links extensively.

The repository content is almost entirely in Chinese. Engineers who do not read Chinese will find the headings and directory names accessible but the explanatory prose and interview preparation materials inaccessible without translation. The code examples and framework documentation it links to are largely in English, but the navigation layer is Chinese.

The interview question bank claims 1,500 or more questions, but the repository provides no information about how those questions are curated, how recently they were updated, or how well they map to current hiring practices at specific companies.

Finally, the README does not document installation or setup for AgentGuide itself as a software project. It is a documentation site deployed at https://adongwanai.github.io/AgentGuide. The README provides no commands for running the site locally.

Comparison with JavaGuide and Maintenance

The README explicitly frames AgentGuide as the AI Agent equivalent of JavaGuide, a well-known Chinese knowledge base for engineers preparing for Java and backend development positions. The structural parallel holds: both repositories organize a learning path, annotate topics with interview relevance, and include interview question banks. The difference is subject matter. JavaGuide covers Java, Spring, databases, and distributed systems; AgentGuide covers LLM application engineering, Agent systems, RAG, and post-training methods.

The last push to AgentGuide was on 2026-09-15, confirming active maintenance at the time of this article. The repository has no GitHub releases and uses continuous MDX publishing rather than versioned releases. The license is not documented in the README or the repository top-level file listing. This gap is worth confirming before using the materials in a commercial context.

Editorial conclusion

AgentGuide is the right choice for engineers in China who want a structured path into AI Agent development or algorithm roles and are willing to work through its Chinese-language content. It is the wrong tool for engineers who need ready-to-run code rather than a curated reading and practice list. Before starting, verify that the algorithm or development track matches your target role; the README decision tree between the two paths is the first document worth reading.

Frequently asked questions

Official sources

1. adongwanai/AgentGuide on GitHub
2. Issues
3. README