Following the release of OpenAI's new Agents SDK, we've seen a lot of interest in multi-agent workflows. Here, we discuss two different approaches for multi-agent systems - swarm and supervisor - and showcase two different LangGraph packages that make it easy to implement these approaches. We show that both can be run locally with Qwen2.5-14b (via Ollama), which excels at tool-calling. We also show LangGraph Studio and LangSmith traces to provide debugging and observability into these systems.
LangGraph-swarm:
https://github.com/langchain-ai/langgraph-swarm-py
LangGraph-supervisor:
https://github.com/langchain-ai/langgraph-supervisor-py
Video notes:
https://mirror-feeling-d80.notion.site/Fully-Local-Multi-Agent-1b5808527b178066bde0ed981b27998c?pvs=4
Chaters:
00:00 - Introduction to Multi-Agent Systems and Open Eye SDK
00:30 - Demo: Flight and Hotel Booking Multi-Agent System
01:10 - Running Locally with Qwen models
02:00 - What is an Agent? (Tool Calling in a Loop)
03:00 - Finding Local Models for Agent Development
04:00 - Berkeley Function Calling Leaderboard and Qwen Models
05:00 - Why Multi-Agent Systems Matter
06:00 - Supervisor vs. Swarm Architecture Explained
07:15 - Trade-offs Between Different Multi-Agent Approaches
08:00 - Building Multi-Agent Systems in a Notebook
09:00 - Understanding the Agent Implementation
10:00 - Setting up the Supervisor Architecture
11:00 - Tracing and Visualization with LangSmith
12:00 - Choosing the Right Local Models for Your Agents
12:45 - Conclusion and Final Thoughts