Microsoft AutoGen

What is Microsoft AutoGen?

Microsoft AutoGen is an open-source Python framework, originally released by Microsoft Research in 2023, for building multi-agent AI applications where agents communicate through structured message passing. Agents run code, call tools, delegate subtasks, and request human input, either autonomously or with supervision checkpoints. As of 2025, AutoGen has accumulated over 45,000 GitHub stars and sees broad use in research and production automation workflows.

AutoGen is now in maintenance mode. Microsoft released Microsoft Agent Framework in 2025 as the enterprise-ready successor, with stable APIs, long-term support, and cross-runtime interoperability via A2A and MCP. Existing AutoGen deployments continue to work; new projects should evaluate Microsoft Agent Framework for production use.

For teams building automation workflows that need role-based agent specialization, CrewAI offers an actively maintained alternative with a similar Python API. For stateful graph-based execution with conditional branching, LangGraph provides more fine-grained control.

How does it work?

AutoGen agents are objects with a name, system message, and model configuration. They communicate by sending messages to each other. The framework handles the conversation loop, tool call dispatch, and termination conditions. Two primary APIs serve different abstraction levels.

AgentChat API

The AgentChat API provides pre-built agent types and a GroupChat orchestrator for rapid prototyping:

• AssistantAgent: LLM-powered agent that reasons and calls tools
• UserProxyAgent: proxies human input or executes code on behalf of the conversation
• GroupChat: coordinates up to 30 agents in round-robin or LLM-selected speaker order
• SocietyOfMindAgent: wraps a multi-agent chat as a single agent callable from an outer loop

Core API

The Core API exposes the message-passing runtime directly. It supports event-driven agent loops, local and distributed execution, and custom runtime implementations. Use Core when you need precise control over message routing, agent lifecycle, or distributed deployment.

Tool and code execution

Capability	How it works
Function calling	Agents declare tools as Python functions; the framework dispatches calls and returns results
Code execution	UserProxyAgent executes Python or shell code in a sandboxed subprocess
Human in the loop	UserProxyAgent pauses execution and waits for terminal input at configurable breakpoints
LLM flexibility	Any OpenAI-compatible endpoint: GPT-4o, Claude, Mistral, Ollama local models

AutoGen Studio

AutoGen Studio is a no-code GUI (install via pip install autogenstudio) for building and testing agent workflows without writing Python. It generates AutoGen configuration JSON that you can export and run in code. Studio is useful for prototyping agent topologies before committing to an implementation.

When should you use it?

Use AutoGen when you need a conversational multi-agent pattern where agents reason about each other’s outputs: coding pipelines (planner to coder to reviewer), research workflows (searcher to summarizer to critic), or iterative content generation. The human-in-the-loop pattern via UserProxyAgent suits supervised automation where a human approves intermediate steps.

For new production deployments requiring long-term support, evaluate Microsoft Agent Framework before committing to AutoGen. For workflows that map naturally to role-specialized teams rather than conversational agents, CrewAI is more actively developed. For complex branching logic that conversational agents handle poorly, LangGraph provides a directed graph execution model with explicit state management.

Frequently asked questions

What replaced AutoGen for enterprise use?

Microsoft Agent Framework (MAF), available at github.com/microsoft/agent-framework, is AutoGen’s enterprise successor. It builds on AutoGen’s architecture with production-ready APIs, cross-runtime support via A2A and MCP, and a long-term maintenance commitment. A migration guide is available in the Microsoft documentation for teams moving existing AutoGen applications to MAF.

How many agents can a GroupChat coordinate?

GroupChat supports up to approximately 30 agents in a single conversation loop. Speaker selection uses either round-robin ordering or an LLM-based selector that picks the next agent based on conversation state. For workflows with more agents or complex routing logic, the Core API’s distributed runtime handles larger topologies; GroupChat covers most research and production use cases within its default limits.

Can AutoGen agents call MCP tools?

AutoGen does not natively implement the MCP client protocol. You can expose MCP tools to AutoGen agents by wrapping them as Python functions that call the MCP tool via HTTP or subprocess, then registering those wrappers as agent tools. Microsoft Agent Framework has native MCP interoperability via the MAF runtime’s MCP transport support.