Authority AI Editorial
AutoGen: Microsoft's multi-agent conversation framework
AutoGen is an open-source framework from Microsoft for building applications where more than one AI agent works on a task, passing messages back and forth instead of running a single prompt-and-response loop. It is built for developers who want agents that can plan, call tools, run code and check each other's work, and it is aimed at people comfortable writing Python (there is also a .NET path for the core runtime). If you already run coding agents day to day and want to script a handoff between a researcher agent and a coder agent, or a debate between two agents before a final answer, AutoGen gives you the pieces to wire that up yourself.
What it does
AutoGen organizes work around agents that exchange messages in a shared conversation. An agent can be a thin wrapper around a language model, a tool executor, or a human proxy that pauses the flow for approval. You define which agents exist, what each one is allowed to do, and how control passes between them, then let the conversation run until the task is done or a stopping condition is hit. The project ships with patterns for two-agent conversations, group chats with several agents, and code-execution agents that run and check generated code before handing results back.
How it is structured
The project is organized in three layers. Core is the event-driven foundation: agents send and receive typed messages, and the runtime can be distributed across processes or machines, with support for more than one programming language. AgentChat sits on top of Core as a Python 3.10-plus API built for quickly assembling common patterns like a two-agent chat or a round-robin group chat, and it is also the layer that AutoGen Studio, a no-code web UI, is built on. Extensions is the connector layer: integrations for OpenAI and Azure OpenAI models, Docker-based code execution, Model Context Protocol (MCP) servers, and gRPC for distributed runtimes. The documentation recommends three entry points depending on how much code you want to write: Studio for no-code prototyping, AgentChat for programmatic prototyping, and Core when you need full control over a distributed, production-grade system.
To install the Python path, pip install -U "autogen-agentchat" "autogen-ext[openai]" gets you AgentChat plus OpenAI support, and you supply your own model API key. The framework itself does not host or serve models.
What to know before adopting it
AutoGen's code is MIT licensed, with its documentation under a Creative Commons Attribution 4.0 license, so you can read, fork and ship on top of it freely. The bigger thing to know is where the project sits inside Microsoft right now. On October 1, 2025, Microsoft announced Microsoft Agent Framework, a new open-source engine that combines ideas from both AutoGen and Semantic Kernel. Microsoft's own announcement states that AutoGen "will remain supported but most investment is now focused on Microsoft Agent Framework," which puts AutoGen in maintenance mode rather than active feature development. If you start a new project today, it is worth reading Microsoft's migration notes, which map AutoGen concepts like GroupChat and GraphFlow onto Agent Framework's Workflow abstraction, and AssistantAgent onto Agent Framework's ChatAgent. AutoGen is still a reasonable choice if you want a mature, well-documented multi-agent conversation model and do not need Microsoft's newest enterprise integrations, but plan for an eventual migration rather than treating it as the long-term target.
Also worth knowing: AutoGen assumes you are comfortable configuring model credentials and, for code-execution agents, a Docker environment to run generated code safely. It is a library and a set of patterns, not a hosted product, so there is no separate cost model beyond whatever model API calls and compute you use.
Where it sits next to other agent tools
AutoGen is squarely in the "orchestration framework" category: it is concerned with how agents talk to each other and to tools, not with editing files in a repository or driving a browser. That puts it next to frameworks like CrewAI and LangGraph rather than next to a coding agent like Claude Code, Codex or OpenHands, which you could in principle wire up as one of AutoGen's agents. If your job is mostly running a single coding agent against your own repository, AutoGen may be more machinery than you need. If you are building a system where several specialized agents need to negotiate, review each other's output, or split a task in a repeatable way, it is one of the most established places to start, and its documentation and migration guides are a useful reference even if you end up on Microsoft Agent Framework later. Separately, if you want ready-made methods to hand any of these agents rather than writing every one yourself, the AI Ultimate Skill Collection is 3,029 skills in 12 packs, each a written procedure your agent follows step by step.