AutoGen
Microsoft's framework for multi-agent conversation systems
- Best for
- Research and complex multi-agent prototypes
- Pricing
- Free and open-source (MIT)
- Stack layer
- AI Stack
Live preview of microsoft.github.io/autogen
Overview
AutoGen was one of the first frameworks to treat agent-to-agent conversation as the unit of computation. Two or more agents talk; one might be a coder, another a critic, another a user proxy that executes code in a sandbox. The "conversation" is the trace of the work.
The strength is research-grade flexibility — almost any multi-agent pattern can be expressed. The cost is that you'll write more glue code than with CrewAI or LangGraph for the same workflow.
AutoGen v0.4 split the project into Core, AgentChat, and Extensions for cleaner separation of concerns.
Key Features
Conversable Agents
Agents that talk to each other to solve tasks
Code Execution
Built-in sandboxed code runner for tool use
Group Chat
Orchestrate N-agent conversations with a manager
Human Proxy
Put a human in the loop as another agent
Modular Core
Use Core, AgentChat, or Extensions independently
Microsoft Backed
Active development from Microsoft Research
Why We Recommend AutoGen
When the workflow doesn't fit a graph or a crew, AutoGen's conversation primitive is the most flexible escape hatch.