This library provides a Python framework for building swarm-style multi-agent systems using LangGraph, enabling specialized agents to dynamically hand off control and context to one another. It's designed for developers creating complex agentic workflows that require collaboration and stateful interactions between distinct AI agents.

How It Works

The core mechanism involves agents created with LangGraph's create_react_agent that incorporate custom create_handoff_tool functions. These tools allow agents to explicitly transfer execution to another agent by name. The create_swarm function orchestrates these agents into a graph, managing the active agent and ensuring continuity. Handoffs can pass the full message history or custom data, and the system supports memory via checkpointers for stateful, multi-turn conversations.

Quick Start & Requirements

• Install: pip install langgraph-swarm langchain-openai
• Requires: OpenAI API key (export OPENAI_API_KEY=)
• Example usage and memory configuration: https://github.com/langchain-ai/langgraph-swarm-py#quickstart

Highlighted Details

• Enables dynamic agent handoffs based on specialization.
• Supports streaming, short-term, and long-term memory.
• Allows customization of handoff tools and agent implementations.
• Built on LangGraph, leveraging its graph-building capabilities.

Maintenance & Community

• Project maintained by langchain-ai.
• Further customization examples and guides are available in the repository.

Licensing & Compatibility

• License: MIT License
• Compatible with commercial use and closed-source applications.

Limitations & Caveats

Custom handoff tools returning Command require agents to have a tool-calling node and shared state schemas that accommodate Command.update fields. Custom agent implementations necessitate careful state mapping between the swarm and individual agent graphs.