Keep a fast AI-assisted team from breaking things

Your team ships fast with agents: 2 engineers now ship what used to take 5, and the org is thrilled. Leadership does not want to slow down.

You see the other side. Approvals arrive within minutes of a pull request opening. Last week an agent-written change to an endpoint another team calls passed review and took down their service for 20 minutes.

How do you manage the risk without killing the speed, and what failure mode worries you that nobody is naming?