Canary deploys: roll out an agent gradually
Route a slice of live runs to a new agent version to prove it in production before full rollout.
Updated Aug 3, 2026
When you change an agent — a new prompt, a different tool — you don't have to switch every run over at once. A canary deploy sends a share of live runs to the new version while the rest stay on the proven one.
Set it up
On a worker's Versions card, pick a candidate version and a percentage. That share of runs is routed to the candidate (deterministically, so a given run always lands on the same version); the rest run on the active version. Compare their outcomes, then promote the candidate to 100% or roll it back to 0%.
Why
It's the safe way to ship a change to an agent that acts in the real world: you get real production signal on the new version at a fraction of the blast radius, with an instant rollback if it underperforms. Canary sits after the evaluation gates — a version still has to pass its golden and adversarial tests before it can take any live traffic at all.