a blast-radius problem

most companies do not have an agent intelligence problem. they seem to have a blast radius problem

agents become deployable when their permissions are tightly scoped, their actions are easy to inspect, and mistakes can be reversed without creating broader damage. that makes draft states, approval gates, audit logs, and rollback systems just as important as model performance

the real production metrics are not only task completion, but how often humans need to intervene, how quickly errors are caught, and how expensive they are to recover from