Learn Claude Code
從 0 到 1 構建 nano Claude Code-like agent，每次只加一個機制
核心模式
所有 AI 程式設計 Agent 共享同一個迴圈：呼叫模型、執行工具、回傳結果。生產級系統會在其上疊加策略、許可權和生命週期層。
while True:
response = client.messages.create(messages=messages, tools=tools)
if response.stop_reason != "tool_use":
break
for tool_call in response.content:
result = execute_tool(tool_call.name, tool_call.input)
messages.append(result)訊息增長
觀察 Agent 迴圈執行時訊息陣列的增長
學習路徑
20 個漸進式課程，從簡單迴圈到完整多 Agent Harness
The Agent Loop
The smallest useful agent is a loop that calls the model, runs tools, and feeds results back.
Tool Use
The loop stays stable while capabilities register into a dispatch table.
Permission
Dangerous actions need a harness decision point before the shell runs.
Hooks
Cross-cutting behavior belongs around the loop, not tangled inside it.
TodoWrite
Explicit plans keep long-running work visible and correctable.
Subagent
Subagents give each subtask a clean message history while preserving the main thread.
Skill Loading
Inject specialized knowledge only when the task actually needs it.
Context Compact
Compression keeps the conversation usable when the context window gets crowded.
Memory
Some facts should survive summarization and future sessions.
System Prompt
The system prompt is a generated product of policy, tools, skills, and context.
Error Recovery
A robust harness classifies failures and decides what kind of retry is worthwhile.
Task System
A task graph turns vague goals into ordered, observable work.
Background Tasks
The agent can keep reasoning while slow work completes elsewhere.
Cron Scheduler
Recurring work should be created by the harness, not remembered by the model.
Agent Teams
Persistent teammates let work continue in parallel without stuffing every thought into one context.
Team Protocols
Multi-agent systems need explicit message contracts, not vibes.
Autonomous Agents
Teammates become useful when they can discover and claim work themselves.
Worktree Isolation
Parallel agents need isolated filesystems as much as isolated conversations.
MCP Tools
External services can become agent tools through a standard discovery and call protocol.
Comprehensive Agent
The final harness is still one loop, now surrounded by the systems that make it production-shaped.
CLAUDE.md
AI doesn't disobey — it guesses when you leave blanks. A written manual removes the guessing.
Verification
Fluent text is not the same as correct output. Define 'done' before the task starts.
Memory
Memory lives in your files, not in AI's head. A three-part handoff note bridges any conversation gap.
Hooks
CLAUDE.md is a guide. A hook is a guard. If 5% failure is unacceptable, enforce it.
Layered Governance
AI ignores the second half of a 200-line rule file. Split into global / task / personal layers.
Integration
Start with a minimum viable harness — manual, checklist, handoff note — and add pieces only when needed.
架構層次
五個正交關注點組合成完整的 Agent