A Chinese-language curriculum for actually understanding AI agents

What it does

Agent Craft is a structured tutorial repository — not a framework, not a library. It’s a 15-module learning path (13 published so far) that builds an AI agent capability stack in Python: raw LLM calls, Function Calling, LangChain, RAG with FAISS and Chroma, LangGraph state machines, MCP servers and clients, and multi-agent orchestration with the Agents SDK and Swarm patterns. Each module is a self-contained directory with commented, runnable code, paired with a long-form CSDN blog post explaining the design decisions.

The interesting bit

The stated philosophy is “don’t reinvent wheels, but don’t stop at framework calls either” — the modules deliberately show how agents think and decide inside LangChain/LangGraph rather than treating the framework as a black box. Module 08, for instance, implements the ReAct loop from scratch as a “white-box” exercise before leaning on the framework.

Key highlights

• Progressive path: Prompt → LLM → LangChain → RAG → LangGraph → MCP → multi-agent → deployment
• Covers both MCP server (FastMCP, Stdio and Streamable HTTP transports) and client (langchain-mcp-adapters) sides
• Advanced topics included: Reranker-based RAG, Human-in-the-Loop, Graph-as-a-Tool, Swarm handoffs
• Module 13 builds a working Streamlit chat product (“intelligent customer-service cockpit”) wired to the Agents SDK
• Has CI, and the content is actively maintained — modules 14–15 (capstone project, deployment with Ollama/LangServe) are still in progress

Caveats

• All teaching material and blog posts are in Chinese; non-Chinese readers will get code but not the explanations
• config.py validates four API keys (DeepSeek, LangSmith, Amap, ChatGPT) on import — all must be non-empty even for modules that don’t use them, which is a small friction point the README itself acknowledges
• Modules 14 and 15 are not yet written, so the “full stack” story currently ends at Streamlit

Verdict

Worth a look if you read Chinese and want a guided, code-first path from “I can call an API” to “I can build and deploy a multi-agent system.” If you’re past the tutorial stage or need English materials, skip it — it’s a curriculum, not a tool.