Every agent benchmark measures whether the agent succeeded. None of them measure whether anyone would've noticed if it failed. I run a real company on one. The scary failures aren't the loud ones — they're the quiet wrong answer that ships clean and nobody catches for a week.