datawhalechina/deepagents-in-action

What this project is

Deep Agents 实战 is a free, Chinese-language online course that teaches you how to build AI agents using the LangChain and LangGraph frameworks. It's published by the DataWhale open-source community and written by 沧海九粟, a LangChain-certified ambassador.

The course has 17 chapters of illustrated written lessons plus ready-to-run experiment templates, with video supplements being added over time. It starts from your very first agent and gradually covers task planning, file systems, sub-agent collaboration, and context management.

Who it's for

Smart but non-expert learners who want to get hands-on with AI agents. The course assumes you'll follow along with experiments, not just read theory. It's in Chinese.

What you'll learn (chapter map)

Preparation — Set up your development environment using a tool called AgentSeek, and install AI-coding-assistant "skills" that give you framework-level guidance while you code.

Version update guide — Explains what changed in Deep Agents v0.7 (lighter default setup, ~65% reduction in base input tokens for simple turns, new file-tool behaviors) and how to migrate from v0.5/v0.6.

Cognitive chapters

• Ch 1: What an "Agent Harness" is and how Deep Agents, LangChain, and LangGraph divide responsibilities
• Ch 2: Build your first agent in 5 minutes

Core chapters

• Ch 3: Virtual file system — how agents save content and intermediate results
• Ch 4: Task planning — teaching agents to break down complex tasks
• Ch 5: Sub-agents — delegating work and isolating context
• Ch 6: Async sub-agents — running multiple subtasks in parallel

Advanced chapters

• Ch 7: Skills — reusable capability packages
• Ch 8: Long-term memory — cross-conversation memory
• Ch 9: Human-in-the-loop — safe human approval flows
• Ch 10: Sandbox execution — running code safely
• Ch 11: Filesystem permissions — declarative read/write boundaries
• Ch 12: MCP — a standard protocol to extend the agent's tool ecosystem

Frontier preview

• Ch 13: Grading rubrics — agents self-iterate against acceptance criteria
• Ch 14: Streaming — watching agent and tool events in real time
• Ch 15: Interpreters — agents orchestrating tools with code
• Ch 16: Dynamic subagents — code-orchestrated multi-agent patterns
• Ch 17: Decision Model Harness — model routing and Auto Mode for agent control

How you follow along

Each chapter has an experiment template. You install AgentSeek, create a template (e.g., agentseek create deepagents/default), configure a .env file, run agentseek doctor to check your environment, then agentseek dev to launch. Models are accessed through SiliconFlow (硅基流动), a cloud model platform; the course's sponsor provides a 16-yuan voucher for new users, and leftover referral credits are pooled to subsidize students who run out of quota.

Where to access it

• Online reading: datawhalechina.github.io/deepagents-in-action/
• Bilibili videos: linked from the README
• Xiaohongshu posts: linked from the README

Tech stack for the site itself

The course website is built with Astro 6, Tailwind CSS 4, and TypeScript. Chapter content lives as Markdown files in a content/ folder; a preprocessing script injects metadata from scripts/chapters.json before building. Local development requires Node.js ≥ 22.12.0.

Licensing

• Course text: CC BY-NC-SA 4.0 (share-alike, non-commercial)
• Website source code: MIT

How to contribute

Pull requests are welcome — fixing typos, improving layout, or discussing content. All contributors are acknowledged in the README's "Special Thanks" section and receive a LangChain official community (China) gift.

Key facts at a glance

Item	Detail
Language	Chinese
Framework taught	LangChain / LangGraph / Deep Agents
Chapters	17 + preparation + version-update guide
Author	沧海九粟 (LangChain certified ambassador)
Model provider	SiliconFlow (硅基流动)
Content license	CC BY-NC-SA 4.0
Code license	MIT
Current version	Deep Agents 0.7.x recommended

Note: The README mentions the course has received 2,000+ GitHub stars and appears on Trendshift, but exact star counts aren't stated in the README itself.