I captured a hundred agent failures a week and fixed three. The bottleneck was curation.

I had full traces and an eval set that was mostly the same bug on repeat. Here is the curation discipline that fixed it.