Pick any step in a finished run, edit its input or the prompt behind it, and play the run forward. Earlier steps reuse cached outputs, so a replay costs only the steps that actually change.

Policies gain a third action. Instead of failing a step, a guardrail can send the output back to the model with the violated rule attached, up to a retry limit you set.

The Go SDK leaves beta with full support for durable steps, context propagation across goroutines, and a net/http middleware that traces inbound requests.

Run your datasets on every pull request with one command. Norvane posts a score summary to the PR and fails the check when a scorer drops below your threshold.

Group any alert by a run attribute such as customer_id or plan. Find the one tenant whose agent loops before it shows up on your invoice.