datawhalechina/agent-learning-hub

This repository is called Agent Learning Hub. It is not a piece of software in the usual sense. The whole project is one long README, written in a mix of English and Chinese, that lays out a curated learning roadmap for people who want to build AI agents instead of just reading random links about them. It is maintained by a Datawhale community member named Chen Sizhou. The README opens with a guide for different kinds of readers. Beginners are told to follow a Learning Todo List in order, ticking items as they go. People who already know how to call large language model APIs are pointed at Stage 2 or Stage 3, which cover the agent loop, tool calling, evaluation, and engineering. People who want to ship projects are sent to a Project Ladder. People who only want references are sent to a Curated Resources section. The roadmap itself is divided into stages. Stage 0 explains what an agent is and how it differs from a chatbot or a fixed workflow. Stage 1 walks through building a minimal agent loop with one model, structured JSON output, and a few tool functions. Stage 2 covers retrieval-augmented generation, memory, and turning search, databases, files, browsers, and code execution into tools. Stage 3 asks the reader to study one modern agent system in depth, with Claude Code, OpenClaw, Hermes, LangGraph, and others suggested. Later stages cover multi-agent coordination, skills and protocols like MCP and A2A, browser and computer-use agents, and evaluation, observability, and safety. The author also gives an opinion on what is worth studying right now. The README puts coding agents, agent harness engineering, personal local-first agents, the Skills and MCP family of protocols, and evaluation at the top. It explicitly recommends against spending most of your time on older role-play multi-agent frameworks. Each stage lists concrete checkboxes, recommended reading from official documentation, and example open-source projects. The output of the whole repository is meant to be a personal learning checklist rather than runnable code.