Learn Claude Code
Overview
12 节渐进式课程，从零构建类 Claude Code 的 agent harness。GitHub 54.7k stars，MIT license。
GitHub: https://github.com/shareAI-lab/learn-claude-code License: MIT | Stars: ~54.7k | Language: Python
Core Philosophy
"Agency comes from the model. Agent Product = Model + Harness."
The code doesn't make an agent smart — the model does. The harness just provides tools, knowledge, observation, action interfaces, and permissions.
What an Agent is NOT
Prompt plumbing "agents" are NOT real agents:
- Drag-and-drop workflow builders
- No-code "AI agent" platforms
- Prompt-chain orchestration libraries
- If-else branches with LLM API calls
These are Rube Goldberg machines — over-engineered, brittle pipelines with an LLM wedged in as a "glorified text-completion node."
Harness Components
Harness = Tools + Knowledge + Observation + Action Interfaces + Permissions
Tools: file I/O, shell, network, database, browser
Knowledge: product docs, domain references, API specs, style guides
Observation: git diff, error logs, browser state, sensor data
Action: CLI commands, API calls, UI interactions
Permissions: sandboxing, approval workflows, trust boundaries
12 Progressive Sessions
| Session | Topic | Motto |
|---|---|---|
| s01 | The Agent Loop | "One loop & Bash is all you need" |
| s02 | Tool Use | "Adding a tool means adding one handler" |
| s03 | TodoWrite | "An agent without a plan drifts" |
| s04 | Subagents | "Break big tasks down; each subtask gets a clean context" |
| s05 | Skills | "Load knowledge when you need it, not upfront" |
| s06 | Context Compact | "Context will fill up; you need a way to make room" |
| s07 | Tasks | "Break big goals into small tasks, order them, persist to disk" |
| s08 | Background Tasks | "Run slow operations in background; the agent keeps thinking" |
| s09 | Agent Teams | "When task is too big for one, delegate to teammates" |
| s10 | Team Protocols | "Teammates need shared communication rules" |
| s11 | Autonomous Agents | "Teammates scan the board and claim tasks themselves" |
| s12 | Worktree + Task Isolation | "Each works in its own directory, no interference" |
Claude Code Architecture
Claude Code = one agent loop + tools + on-demand skill loading + context compression + subagent spawning + task system with dependency graph + team coordination with async mailboxes + worktree isolation + permission governance
Related
- nanobot — 用 Python 实现 agent harness 的项目
- hermes-agent — 生产级 agent 框架