四周时间，零框架手写 ReAct → MCP 工具治理 → 弹性容错 → LangGraph 状态机 → Multi-Agent 协作。 完整学习看板 → tallmessiwu.github.io/agent-building
git clone --recurse-submodules https://github.com/TallMessiWu/agent-building.git
uv sync
# 日常练习（按时间线推进）
uv run python exercises/w1d1-function-calling/agent.py
uv run python exercises/w1d2-mcp-server/agent.py
uv run python exercises/w1d3-resilient-agent/agent.py
uv run python exercises/w2d1-langgraph/agent.py
uv run python exercises/w2d2-langgraph-hitl/agent.py
uv run python exercises/w2d3-autogen/agent.py
uv run python exercises/w2d4-crewai/agent.py
uv run python exercises/w2d5-rag-agent/agent.py --self-test
uv run python exercises/w3d1-reflexion/agent.py --self-test
uv run python exercises/w3d2-plan-execute/agent.py --self-test
uv run python exercises/w3d3-langgraph-supervisor/agent.py --self-test
uv run python exercises/w3d4-agent-evaluation/agent.py --self-test
# 交互式 Notebook
uv run jupyter notebook exercises/w1d1-function-calling/agent.ipynbagent-building/
├── README.md
├── CLAUDE.md
├── Agent_4周学习计划.xlsx
├── docs/index.html # → GitHub Pages 学习看板
│
├── exercises/ # 每日编程练习（12 天）
│ ├── w1d1-function-calling/ # Day 1: Function Calling
│ ├── w1d2-mcp-server/ # Day 2: MCP 协议
│ ├── w1d3-resilient-agent/ # Day 3: 重试/熔断/降级
│ ├── w2d1-langgraph/ # Day 4: StateGraph + Checkpoint
│ ├── w2d2-langgraph-hitl/ # Day 5: HITL + Subgraph
│ ├── w2d3-autogen/ # Day 6: AutoGen Multi-Agent
│ ├── w2d4-crewai/ # Day 7: CrewAI Role-based Agent
│ ├── w2d5-rag-agent/ # Day 8: Chroma RAG + LangGraph
│ ├── w3d1-reflexion/ # Day 9: Reflexion + Self-Refine
│ ├── w3d2-plan-execute/ # Day 10: Plan-and-Execute + DAG
│ ├── w3d3-langgraph-supervisor/ # Day 11: LangGraph Supervisor
│ └── w3d4-agent-evaluation/ # Day 12: Agent Evaluation
│
├── react-agent/ # 📦 项目 1: 手写 ReAct Agent (submodule)
├── long-term-memory-research-agent/ # 📦 项目 2: 长期记忆研究助手 (submodule)
└── multi-agent-collab/ # 📦 项目 3: Multi-Agent 协作写作
📅 完整学习时间线、41 道八股题库、必读清单、项目追踪、模拟面试反馈 → 网站版学习看板
| 亮点 | 位置 | 说明 |
|---|---|---|
| 受限 eval 沙箱 | w1d1/agent.py |
__builtins__ 置空 + 白名单，面试高频考点 |
| MCP 协议落地 | w1d2/agent.py |
Anthropic 主导的 Agent 工具标准协议 |
| Circuit Breaker | w1d3/agent.py |
三态状态机熔断器，分布式系统经典模式 |
| 指数退避 + 抖动 | w1d3/agent.py |
避免雷鸣羊群效应，生产级重试 |
| 混沌工程 | w1d3/agent.py |
主动注入故障验证弹性 |
| LangGraph StateGraph | w2d1/agent.py |
图结构控制流 + SQLite Checkpoint 持久化 |
| HITL 断点恢复 | w2d2/agent.py |
零侵入人工审批 + 子图横切关注点 |
| AutoGen Multi-Agent | w2d3/agent.py |
LLM 动态选择发言人 + 三框架完整对比 |
| CrewAI Role-based | w2d4/agent.py |
role/goal/backstory + Task 顺序协作 |
| Chroma RAG | w2d5/agent.py |
本地向量库 + 混合检索 + LangGraph 工具调用 |
| Reflexion | w3d1/agent.py |
generator/critic/refiner 闭环 + 反思记忆 |
| Plan-and-Execute | w3d2/agent.py |
任务计划 DAG + 可并行批次调度 |
| LangGraph Supervisor | w3d3/agent.py |
supervisor 路由 + Researcher/Writer/Critic 修订闭环 |
| Agent Evaluation | w3d4/agent.py |
成功率 + 轨迹评估 + LLM-as-judge 风格 rubric + trace |
Python 3.14 · OpenAI SDK (DeepSeek API) · MCP Python SDK · LangGraph · AutoGen · Chroma · uv