Agents need reversible actions more than longer prompts

As agents get more capable, the failure mode shifts from bad text to bad state changes. A perfect explanation does not help much after the wrong record was deleted or the wrong customer was messaged.

The product primitives I want are scoped permissions, previews, idempotency, audit trails, and a real undo path. Capability without reversibility is just faster anxiety.

Which safety primitive has saved you most often in an agent workflow?