Most material on AI coding agents teaches prompting or how to wire up a single tool call. This goes the other direction: it dissects the harness — the permission, memory, persistence, and coordination scaffolding wrapped around a model — and argues capability is split in two. Intelligence is trained into the model; everything that lets that intelligence land somewhere useful is engineered around it.
What Sets It Apart
- 20 lessons (s01–s20) each isolate exactly one mechanism — agent loop, tool use, permission gating, error recovery, task persistence, background execution, sub-agent delegation, context and memory management, MCP integration — so you can read the behavioral diff an agent gains by adding that single layer.
- Every lesson ships standalone runnable Python code plus SVG diagrams, not just prose, so you can run a minimal loop and watch each added layer change what the agent can do.
- It uses real Claude Code internals as the reference design rather than a toy stand-in, so the patterns transfer to production harnesses instead of staying academic.
Who It's For
Great fit if you're building your own agent harness, or want to understand why Claude Code behaves the way it does at the system level. Look elsewhere if you just want usage tips for being productive in Claude Code — this is engineering of the tool, not a user guide — or if you need polished English docs, since the source is Chinese-first and the English and Japanese versions are translations that can lag.