CrewAI is an open-source framework for building multi-agent systems around a single organising metaphor: the crew. You define agents as role-holders — a researcher, an analyst, a writer — each with a goal, a backstory that shapes its behaviour, and tools it may use; you define tasks and how they flow; the framework runs the collaboration. The pitch is expressiveness: where graph-based frameworks make you draw control flow, CrewAI lets you *cast* it, and a working multi-agent prototype emerges in an afternoon from a description a non-specialist can read.
The metaphor is the framework's genuine contribution and its known limit. It buys fast onboarding, readable system definitions, and a natural fit for workflows that genuinely look like teams handing work along — research-then-write, draft-then-review, gather-then-decide. It strains where production agents always strain: precise control over branching and failure, durable state across restarts, and audit-grade answers to "what exactly happened in this run." The CrewAI–LangGraph comparison maps the trade in full; the common journey — prototype in the metaphor, productionise in the graph — is common because both halves are true.
The governance reading of a crew
A crew is several agents, and the multi-agent rules apply without discount: per-agent identities rather than one shared token, traced hand-offs, and injection treated as something that propagates — content that steers the researcher can reach the writer's tools two steps later. Govern the crew as one registered unit with several actors inside it, and put the autonomy dial on its external writes. The framework's accessibility cuts both ways here, as accessibility always does: crews are easy to stand up, which means crews appear without ceremony, which means the registry rule — no register, no run — is doing more work, not less, in a CrewAI shop.
As with every framework page in this library: version specifics, enterprise offerings, and feature lists move monthly and are deliberately absent — verify current state against the project's own documentation. The landscape page holds the durable advice: the framework is the most replaceable layer, so keep tools, evaluations, and audit requirements portable.