Deploy to production

Use a separate workspace

Production data and experiments do not belong together. A second workspace is the boundary; a naming convention is not.

Pin versions

Runs record the agent version and ontology version they used. Pin the version your production traffic uses and promote deliberately.

Set limits before traffic

Spend limits, concurrency, and approval steps are cheaper to set now than to add after an incident.

Watch the first week

Read traces daily for the first week, not just failures. The runs that succeeded for the wrong reason are the ones worth catching early.

Have a way back

Promoting a version is reversible. Know which version you would roll back to, and confirm it still passes the case set.