ReplayOpen source coming soon
Model upgrades
start with a replay.
Run your recorded agent workflows with a new model. See how its actions change before you deploy.
Replay in practice · Synthetic example
A changed key can issue a second refund.
A refund times out. The candidate model retries with a fresh idempotency key. Replay flags the changed write and stops at the unmatched call.
EnvironmentsIn development
Build the capabilities
your models need.
Repeatable tasks, training graders, and sealed evaluations for building specific model capabilities.