\r \r

LangGraph for Agents\r

\r

When to use\r

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

How to refer\r

Integrated Reference Examples\r

Read the examples in "./references/" to understand common patterns. \r Start with "./references/README.md" for an overview, then read the target file, it will show more details.\r \r !Important: \r To build an agent, API_KEY credentials is necessary. \r This is user privacy, please do not hard-code it, just hold a placeholder, e.g. API_KEY=your-api-key, and let the user manage the actual keys.\r \r

External Resources\r

[Search]\r If the "search" tool is available, you can refine the query keywords and execute the search.\r \r [Browse]\r If the "browse" tool is available, you can visit the following three websites:\r

• LangGraph Official GitHub Repository (https://github.com/langchain-ai/langgraph)\r
• LangGraph Official Documentation (https://docs.langchain.com/oss/python/langgraph/overview)\r
• LangChain Official Documentation (https://docs.langchain.com/oss/python/langchain/overview)\r \r [Fetch]\r If the "fetch" tool is available, you can retrieve content from the following URL:\r
• Context-7 LangGraph (https://context7.com/websites/langchain_oss_python_langgraph/llms.txt?tokens=10000)\r You may adjust the number of tokens by modifying the tokens parameter in the URL. The default value is 10,000.\r \r

Project Structure\r

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

├── app/                      \r
│   ├── api/        # API endpoints\r
│   ├── backend/    # LangGraph/LangChain logic\r
│   └── frontend/   # User interface\r
├── .env.example\r
├── requirements.txt\r
└── README.md\r
```\r
\r
## Process for Agent System Design\r
### Step 1: Determine System Level\r
- Single-Agent System: Focus on the internal structure of one agent.\r
- Multi-Agent System: Focus on collaboration and communication between multiple agents.\r
\r
### Step 2: Choose Framework\r
- LangGraph: Best for stateful, complex workflows.\r
- LangChain: Best for standard agent patterns based on tool calling.\r
\r
### Step 3: Design Specific Implementation\r
#### For Single-Agent Systems:\r
- With LangGraph: Build a workflow with several nodes, or implement a ReAct Agent with manual tool_node.\r
- With LangChain: Build a ReAct Agent by `create_agent` API.\r
\r
#### For Multi-Agent Systems:\r
- With LangGraph:\r
  - Option 1: Treat each node as an independent agent, connecting them via the Graph API.\r
  - Option 2: Encapsulate a multi-node workflow as a single agent, calling other agents as tools.\r
\r
- With LangChain:\r
  - Create a main ReAct Agent and encapsulate other agents as tools for collaboration.\r
\r
## Build Philosophy\r
- Prefer Native: Check if a tool or integration already exists in LangChain before custom building.\r
- Single File First: Keep core logic in one file initially to simplify debugging.\r
- Clean Code: Provide only essential comments and use clear, descriptive variable names.\r
- Real Data: Use actual API URLs and schemas whenever possible.\r