What it does

Agent-Learning-Hub is a single README that functions as a self-study curriculum for building reliable AI agents. It stages learning from “what even is an agent” through to evaluation, observability, and safety, with explicit checkboxes and expected outputs at each stage. The maintainer, a Datawhale member, updates it to track where the field is actually going rather than where the hype was six months ago.

The interesting bit

The roadmap is deliberately opinionated about what not to learn. It warns against over-investing in “crew/role-play frameworks” that have become template libraries, and instead steers learners toward harness engineering, skills packaging, and protocols like MCP and A2A. Think of it as a curriculum designer who actually reads release notes.

Key highlights

• Seven staged levels, from minimal agent loop to browser-use agents and safety auditing, each with concrete deliverables (e.g., “a 50-150 line minimal agent”, “a debuggable harness demo with failure records”).
• Curated project references at each stage with explicit pedagogical rationale—why GPT Researcher fits Stage 2, why claw0 fits Stage 3.
• Explicit anti-patterns: when to skip agents entirely (“task predictable, workflow stable, ordinary script sufficient”), when single-agent beats multi-agent, and why demo-grade work fails without eval and trace.
• Heavy emphasis on harness internals—tool registries, permission gates, context compaction, session stores—rather than framework API calls.
• Bilingual resources, with Chinese context and English primary sources; links to Claude Code, OpenClaw, Hermes, and emerging protocols rather than older orchestration frameworks.

Caveats

• The repository is literally one README; there is no code, no automation, no issue tracking beyond the checklist format. It is a reading list with homework, not a framework.
• Some sections reference protocols (A2A, ACP) and projects that are themselves early or evolving; the roadmap’s value depends on continued curation.
• The README was truncated in the provided source, so the full depth of Stage 7 and any later stages or project ladder details are not visible here.

Verdict

Worth bookmarking if you are a developer who learns by building and wants to avoid the agent equivalent of “hello world in 47 frameworks.” Skip it if you need a drop-in library or are already shipping production harnesses and just want protocol specs raw.