// the find
microsoft/autogen
A programming framework for agentic AI
AutoGen is Microsoft Research's multi-agent orchestration framework for Python (and .NET), letting you wire up agents that call tools, talk to each other, and run workflows. It's what kicked off the current wave of multi-agent frameworks. As of early 2026, it's in maintenance mode — Microsoft has shipped a successor called Microsoft Agent Framework and is steering everyone there.
The layered architecture (Core → AgentChat → Extensions) is well thought out: you can drop in at whatever abstraction level you need without fighting the framework. MCP tool support is first-class and works cleanly with the Workbench API. The extension ecosystem (OpenAI, Azure, Anthropic, Gemini, Ollama, SemanticKernel) means you can swap model backends without rewriting agent logic. AutoGen Studio gives non-engineers a way to prototype workflows that actually produces runnable code, not just a demo.
It's in maintenance mode — new features are dead, bug fixes are community-best-effort, and Microsoft is actively pushing you toward a different repo. The v0.2 → v0.4 migration was painful enough that they had to write a dedicated guide, which is a sign the API wasn't stable. State persistence in long-running agents is still underbaked: there's no built-in durable state store, so a crashed agent loses its conversation history unless you build that yourself. The .NET side lags the Python side noticeably — fewer extensions, less documentation, and clearly not the team's priority.