Routing, fallbacks and worker performance
Choose workers using recorded outcomes, token usage, speed, workload and eligibility.
A station can use a fixed worker or choose among eligible candidates. Dynamic routing uses recorded evidence and configured rules; it does not ask a model to choose the worker.
Choose an objective
Candidate selection can account for quality, tokens, speed, workload and clearance, including a token ceiling. Set the station’s routing objective and candidate list. Eligibility and permissions still limit which worker may receive a card.
Fallback candidates provide another attempt when the selected worker fails or cannot proceed. Retry limits and chain limits still bound the total work. An A/B experiment can compare workers over recorded runs; a person decides whether to adopt the result.
Read performance with its sample
Review first-pass rate, duration, token usage and reopened work at the station. Outcome records can add reverts, overrides and production issues, with breakdowns by work category or repository where data exists.
Small samples are shown as insufficient evidence. A ranking from one kind of task should not be assumed to apply to another. These are operational records from your workspaces, not marketplace ratings or certification.
See optimization and outcomes for factory-level recommendations and policies for the limits that routing must respect.