AutoGen is Microsoft Research's multi-agent chat framework with three core actors, UserProxyAgent, AssistantAgent, GroupChatManager, and a turn-selection policy instead of a graph.
Picture a small meeting where one person takes notes, several specialists answer questions, and a moderator decides who speaks next. The note-taker also has hands and can run errands the group asks for. The specialists only talk. The moderator listens to the conversation and picks the next speaker based on what just got said. The meeting keeps going until somebody says we are done. AutoGen sets up exactly this kind of meeting, except the participants are language-model bots and the errands are things like running a snippet of code or fetching a file.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
AutoGen is one of three or four frameworks any senior LLM engineer is expected to be able to place on the map in 2026. Knowing what it is, and what it is not, separates candidates who have built something in the multi-agent space from those who have only read about it.
This deep dive covers the framework's identity (who maintains it, what version is current), the three core agent types and how they compose, the v0.4 layered architecture, and where AutoGen sits in the wider framework landscape compared to LangGraph and CrewAI.
Origin, maintenance, and current version
AutoGen is an open-source multi-agent framework from Microsoft Research, first released in 2023. The project is maintained on GitHub under the microsoft/autogen organization with a sizeable community around it. The current major version line is 0.4, which rewrote the framework around a layered architecture; the older 0.2 line is on extended maintenance.
Why the version matters
The 0.2-to-0.4 transition is one of the larger framework rewrites in the LLM space. Agent definitions, runtime APIs, and the multi-agent patterns all changed. Code samples from 2023 tutorials usually target 0.2 and do not run on 0.4 without modification. Interviewers who ask about AutoGen are usually asking about the 0.4 shape.
Where it sits in Microsoft's stack
AutoGen is not the only multi-agent project at Microsoft. Semantic Kernel covers a different abstraction (a kernel of plugins and planners) and is more enterprise-integration focused. The two coexist and target overlapping audiences; for pure multi-agent conversation workloads, AutoGen is the project most teams reach for.
Situations where this technique stops working.
2–4 min · Everything important, quickly.
Real products, models, and research that use this idea.
- Microsoft Research's AutoGen 0.4 documentation describes the Core / AgentChat / Extensions layering used in production deployments.
- Microsoft's Magentic-One reference agent is built on AutoGen and uses GroupChat for multi-modal task orchestration.
What an interviewer would ask next. Try answering before peeking at the approach.
QHow does AutoGen v0.4's actor-model Core enable cross-process agent deployment?
Core treats each agent as an actor with a typed inbox; messages cross process boundaries via the runtime's transport; agent definitions stay unchanged, only the runtime selection changes between in-process and distributed.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Treating AutoGen as a graph framework like LangGraph and being surprised when the next speaker is chosen by a policy at runtime instead of an explicit edge.
60 second bullets to scan on the way to the call.
The three core AutoGen agent types and what each is for
What GroupChatManager does and how speaker selection works
Primary sources. Browse if you want the original framing.
Same topic, related formats. Practice these next.