AutoGen

Agentic AIFrameworks and SDKsPublished By Simon Budziak

AutoGen is Microsoft Research's open-source framework for building multi-agent applications where agents solve tasks by talking to each other in structured conversation. Development has moved on to Microsoft Agent Framework, its official successor, so most teams now meet AutoGen through its ideas rather than by starting new work on it.

What made AutoGen’s approach different?

Where many frameworks chain tool calls in a fixed order, AutoGen treated coordination itself as conversation: agents exchanged natural-language messages until a task converged, any participant could be backed by a model, a tool, or a person, and a human could join the loop as just another speaker. That message-passing design made it the reference implementation for early multi-agent systems, and it normalized keeping a person reachable inside the loop instead of bolting approval steps on afterwards.

Should you start a new project with AutoGen?

For production work, no. Microsoft folded AutoGen together with Semantic Kernel into Microsoft Agent Framework and positions it as the direct successor, while the original research line continues in community forks. The ideas outlived the codebase: role-based crews in CrewAI and explicit state graphs in LangGraph both answer the same question AutoGen asked first, which is how independent agents should coordinate. What you adopt from AutoGen today is the design pattern, not the package.

Frequently asked questions

Is AutoGen still maintained?

The original library receives upkeep only, while active development went into Microsoft Agent Framework, the successor that combines AutoGen's agent abstractions with Semantic Kernel's enterprise features such as state management and telemetry. Existing codebases keep working, but new projects are pointed at the successor.

What is the difference between AutoGen and CrewAI?

AutoGen coordinated agents through relatively free-form peer-to-peer conversation, while CrewAI assigns each agent an explicit role and goal inside a managed crew. Both target multi-agent work; they differ mainly in how much structure they impose on the coordination.

Summarize this page with

See this working in a system we built