LangGraph Tutorial · Course Guide

本课程共 14 章，基于 2026-09-17 核验的 Python 官方文档编写，是原创教学整理而非官方文档的完整镜像。

LangGraph provides explicit state and execution control; Deep Agents adds filesystem tools, context management, and delegation on top of it. Learn the mechanisms here, then combine them into production services in Agent Engineering in Production.

Prerequisites: Python functions, type annotations, virtual environments. Python 3.11+ recommended. Pure-graph examples do not call a model; model examples need your own tool-calling model service (any OpenAI-compatible endpoint) and may incur cost. All credentials in the code come from environment variables.

环境与版本边界

python3 -m venv .venv
source .venv/bin/activate
python -m pip install -U langgraph
python -m pip freeze > requirements.lock.txt

LangGraph v2 streaming requires 1.1+; current docs also recommend the event stream introduced in 1.2. Deep Agents docs describe 0.7 behavior: planning must be enabled explicitly. Do not mix the website doc version with your installed version; check with python -m pip show langgraph deepagents before running and keep the lock file.

章节路线

1. Run a graph without a model
2. State, reducers, and concurrent updates
3. Conditional routing and bounded loops
4. Messages and the tool-calling loop
5. Checkpoints, threads, and cross-thread memory
6. Human approval, interrupts, and resume
7. Streaming and frontend event contracts
8. Subgraphs and explicit boundaries
9. Functional API and task replay
10. Fault tolerance, retries, and idempotency
11. Test transitions and regression boundaries
12. Capstone: evidence-grounded local QA
13. Capstone 2: multi-agent ticket routing
14. Capstone 3: fault tolerance, approval, and resume

• 官方 LangGraph 概览
• 官方 Deep Agents 概览