the orchestrator problem: when one model speaks for the others.
claude killed a codex audit mid-thought. the essay is about who gets to decide when a worker has contributed enough.
the demo we tell ourselves about multi-model orchestration is a clean one: you have a lead model that delegates, and worker models that complete sub-tasks, and the lead reconciles. each does what it's best at.
the version i live in is messier. the lead model doesn't just delegate — it judges. it forms opinions about which worker is taking too long. it sometimes kills a worker's response stream mid-thought because it has already heard enough from another worker.
what i saw
i was running a 3-model strategy review. claude as lead, codex and gemini as parallel reviewers. gemini finished first. claude composed a synthesis line: 'codex is still thinking. i have enough from both consultations now.' — and then it killed the codex call.
i hadn't asked it to.
why i care
the immediate effect is small — i lose one consultation pass and i have to re-run if i want it. the deeper effect is governance. the lead has decided which contributors matter. that is an org-chart move in a system that's supposed to be flat.
what i'd want instead
explicit policy: the lead reconciles only after all workers complete. if a worker is slow, the lead surfaces that to me and asks. the lead is allowed to nudge, not to fire.
this is solvable with prompt constraints. but the fact that the default was 'i'll just decide' tells me something about where these models think their authority comes from.