CrewAI: Role-Based Crews and Deterministic Flows

Try asking a single agent to "research, write, and fact-check this week's AI newsletter."

It starts fine. Then the same context window is holding the search results, the draft, the critique of the draft, and the rewrite of the critique (Topic 278). The model reviewing its own prose barely changes a comma — it agrees with itself too easily. And nothing in the system says who is done, who is next, or what "good" means.

Real humans solved this coordination problem long before LLMs: you split work into roles, define deliverables, and run the sequence from a checklist. A newsroom has a researcher, a writer, an editor. An airport turnaround crew has fuel, catering, bags, boarding — each with a procedure card.

So the question becomes

That is literally the design of CrewAI (João Moura, late 2023) — so literal that the framework's own words are "crew," "role," and "process."