Deep Agents 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。

Deep Agents is an agent harness built on LangGraph: built-in filesystem tools, skills, memory, subagent delegation, and human-in-the-loop approval. The course goes from your first agent to production-grade composition (approval + idempotency + sandboxes).

Prerequisite: read LangGraph Tutorial chapters 1, 4, and 6 (graphs, tool calling, interrupts). Model examples need your own tool-calling model service (any OpenAI-compatible endpoint); credentials come from environment variables.

环境与版本边界

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

Since Deep Agents 0.7, task planning (write_todos) is opt-in; the delete file tool and the filesystem tool allowlist also require 0.7+. The version verified locally is 0.7.15. With pip show deepagents below 0.7, planning and some file-tool behavior differ.

章节路线

1. Your first deep agent and its file tools
2. Model routing and provider strings
3. Filesystem permissions and path boundaries
4. Sandbox execution and the shell tool
5. Skills and progressive disclosure
6. Memory and AGENTS.md
7. Context compression and prompt caching
8. Task planning and write_todos
9. Subagents and context isolation
10. Human-in-the-loop and conditional interrupts
11. Streaming and subagent events
12. Capstone: files + skills + delegation + approval
13. Capstone 2: read-only codebase audit assistant
14. Capstone 3: planning, delegation, and approval research pipeline

• 官方 Deep Agents 概览
• 官方 LangGraph 概览