datawhalechina/Agent-Learning-Hub

What it is:

Agent-Learning-Hub is a curated, README-only learning roadmap for people who want to build practical, reliable AI agents. It’s maintained by a Datawhale member (陈思州) and deliberately avoids being a random link collection — instead it’s organized as an actionable todo list.

Core philosophy:

The project argues that the field has shifted. Rather than old-style role-play / multi-agent frameworks (like CrewAI-style “crews”), it recommends focusing on:

1. Coding agents (Claude Code, Codex-style) — best real-world agent engineering samples
2. Agent harness engineering — tool protocols, permissions, state, feedback, replay, CI, evals
3. Personal always-on agents (OpenClaw, Hermes-style) — long-running, local-first, cross-app
4. Protocols — Skills, MCP (tools), A2A (agent-to-agent), ACP (host apps)
5. Evaluation & safety — without evals, traces, and permission boundaries, an agent is just a demo

Structure:

• How To Use — four entry points depending on your level (beginner, already know LLM apps, want to build projects, or just want resources)
• Learning Todo List — 9 stages from Stage 0 (“Understand What An Agent Is”) to Stage 8 (“Ship A Real Agent”), each with checkboxes, recommended readings, open-source project references, and a concrete deliverable
• Project Ladder — 11 progressively harder projects, from Calculator Agent to Production Harness
• Curated Resources — official docs/blogs, project map, skills/protocols tooling, modern agent systems, legacy frameworks to avoid or skim, papers, and GitHub repos

Key viewpoints:

• Don’t over-invest in old CrewAI/AutoGen-style role-play frameworks — understand them but don’t make them the main path
• An agent’s capability largely comes from its harness, not just the model
• Distinguish between a tool (callable interface), a skill (reusable procedure knowledge), MCP (external tool/data connection), and A2A (agent-to-agent)
• Evaluation and safety are non-negotiable for production agents

Target audience:

• Beginners who want step-by-step progression
• LLM app developers looking to level up from basic API calls to full agent systems
• Builders who want concrete projects to practice on

Note on content details: The README includes specific checklists (e.g., Stage 1: build a 50-150 line minimal agent loop; Stage 5: write a SKILL.md with name, description, steps, acceptance criteria; Stage 7: produce an eval table with at least 20 tasks). It also maintains a curated list of papers (ReAct, Toolformer, Reflexion, Generative Agents, Voyager, AgentBench, WebArena, SWE-bench, GAIA, OSWorld, τ-bench, etc.) and modern agent systems to study.

Tone: Practical, opinionated, and career-oriented — it explicitly tells readers what to invest time in and what to skip, rather than being a neutral aggregation.