Orchestrator-worker reliability

Orchestrator-worker reliability measures whether a fan-out topology, one hub dispatching a pool of workers, keeps a single worker’s faulty return from reaching the siblings the hub coordinates. Its blast radius approaches the fan-out width in the worst case: when the orchestrator redistributes one worker’s stale or poisoned return as a subtask input to a new round, that return can corrupt every sibling the hub dispatches in a single hop.

This is what separates fan-out from debate and consensus, which also run many workers at once. In debate the workers see each other’s output and a peer can catch a wrong answer; in consensus an independent majority can outvote a lone error. Under an orchestrator the workers never cross-check, so the hub is the only thing that can stop a bad return before it spreads.

So the reliability of the pattern turns on a single gate: whether the orchestrator validates a return before it merges or re-dispatches it. Validated, the fault stays with the worker that produced it; forwarded unchecked, the fan-out width becomes the damage. That gate is the variable cascade resistance scores across topologies, and it fixes where the fan-out pattern sits among the multi-agent failure modes.