Canary releases for agents
Send five percent of traffic to a new version and let the evals watch it.
T
Maya Brandt
6 min
A prompt change can look perfect on your test set and still misbehave on real input. A canary gives you real input with a small blast radius.
Pick the first slice
Five percent is enough for most teams. Choose by customer, not by request, so one person does not see two behaviours.
Watch the same gates
Run the production evals on canary traffic. If a gate trips, traffic returns to the stable version automatically.
