LangGraph for Agents

When to use

• Use this skill when the user asks to build agents or multi-agent systems using LangGraph/LangChain.

How to refer

Integrated Reference Examples

Read the examples in ./references/ to understand common patterns. Start with ./references/README.md for an overview, then read the target file, it will show more details.

External Resources

[Search] If the "search" tool is available, you can refine the query keywords and execute the search.

[Browse] If the "browse" tool is available, you can visit the following three websites:

• LangGraph Official GitHub Repository (https://github.com/langchain-ai/langgraph)
• LangGraph Official Documentation (https://docs.langchain.com/oss/python/langgraph/overview)
• LangChain Official Documentation (https://docs.langchain.com/oss/python/langchain/overview)

[Fetch] If the "fetch" tool is available, you can retrieve content from the following URL:

• Context-7 LangGraph (https://context7.com/websites/langchain_oss_python_langgraph/llms.txt?topic={}&tokens={})

The topic and tokens query parameters can be adjusted based on your needs:

• topic: Filter the documentation to a specific subject area (e.g., interrupt, streaming, checkpoints, tools).
• tokens: Control the response length (e.g., 10000 for detailed docs, 5000 for concise output).
• Example: To fetch ~10k tokens of LangGraph docs focused on the interrupt feature: https://context7.com/websites/langchain_oss_python_langgraph/llms.txt?topic=interrupt&tokens=10000

Project Structure

For demos or tests, use a single .py file. For production-grade applications, use:

├── app/                      
│   ├── api/        # API endpoints
│   ├── backend/    # LangGraph/LangChain logic
│   └── frontend/   # User interface
├── .env.example
├── requirements.txt
└── README.md

Process for Agent System Design

Step 1: Determine System Level

• Single-Agent System: Focus on the internal structure of one agent.
• Multi-Agent System: Focus on collaboration and communication between multiple agents.

Step 2: Choose Framework

• LangGraph: Best for stateful, complex workflows.
• LangChain: Best for standard agent patterns based on tool calling.

Step 3: Design Specific Implementation

For Single-Agent Systems:

• With LangGraph: Build a workflow with several nodes, or implement a ReAct Agent with manual tool_node.
• With LangChain: Build a ReAct Agent by create_agent API.

For Multi-Agent Systems:

• With LangGraph:

  • Option 1: Treat each node as an independent agent, connecting them via the Graph API.
  • Option 2: Encapsulate a multi-node workflow as a single agent, calling other agents as tools.

• With LangChain:

  • Create a main ReAct Agent and encapsulate other agents as tools for collaboration.

Build Philosophy

• Prefer Native: Check if a tool or integration already exists in LangChain before custom building.
• Single File First: Keep core logic in one file initially to simplify debugging.
• Clean Code: Provide only essential comments and use clear, descriptive variable names.
• Real Data: Use actual API URLs and schemas whenever possible.