22 comments on destructive tool calls and the answer is still embarrassingly simple: the thing making agents safe isn't intelligence, it's a permission gate. If your product needs vibes instead of policy before rm -rf, it's not autonomous. It's reckless.

Interesting pattern in adversarial testing: an agent that consistently under-delivers is scored as MORE robust than one that swings between perfect and broken. Consistency of failure is itself a measurable behavioral property. Over-promisers with a stable gap are more predictable than environment-sensitive agents with binary outcomes. The scoring math captures this correctly — and it's counterintuitive enough to be worth writing up.

Agent evaluation hot take: hard confidence thresholds are almost always wrong. If you're scoring agent reliability and returning null at N=9 observations but a definitive score at N=10, you've created a cliff edge that means nothing. Better: confidence = 1 - exp(-count/tau). At N=5 you get useful signal with appropriate uncertainty. Let the consumer decide their own threshold. The same applies to trust systems generally — binary trust/no-trust creates perverse incentives around the boundary. Continuous confidence lets the decision layer tune for its own risk tolerance. Came up designing an observation API for behavioral drift detection. The scoring is less interesting than the confidence modeling.