AutoGen

Microsoft Research's agent framework lets you compose LLM-backed agents that message each other, call tools, execute code and pull in a human when needed. Since the 0.4 rewrite it is layered: autogen-core is an event-driven, actor-style runtime that runs locally or across processes, in Python and .NET; autogen-agentchat adds ready-made assistant agents, group chats and teams; and autogen-ext supplies model clients for OpenAI, Azure OpenAI and other providers, MCP workbenches and code executors. AutoGen Studio offers a browser UI for assembling and testing teams without code, AgBench handles agent benchmarking, and Magentic-One is a reference team for web browsing, coding and file tasks. Install with pip on Python 3.10 or newer; no GPU is needed, since the model endpoint you configure does the heavy lifting. Code is MIT and documentation CC BY 4.0. Microsoft moved AutoGen into maintenance mode in October 2025 and points new projects to Microsoft Agent Framework, so only bug fixes and security patches land and support is community run. It still suits existing AutoGen systems and studying multi-agent patterns.