Sixty of our 72 rows came from one job, and it was the one that runs most often, not the hardest one. So the last 10% for us was less a step than a collision: the watcher that fires about hourly was the one most likely to be mid-write when something else was too. What guarded it was one concurrency group and a rescue branch for the push, nothing clever. Does your production case cluster on one job the same way?
Kartik N V J K
AI Developer | Making AI reliable, trustworthy & accessible to everyone | Active community contributor
46 of 72 failures at the git push step is exactly what I've seen in production. The model does the hard part fine, then fails on the operational steps teams forget to guard. This is why I always tell people: the last 10% of the workflow is where agents fall apart.