Capping GPT-6 Astra agent retries made the system more reliable
The author gave each Astra agent task only 3 retries, logging the blocker and moving on once exceeded, and came back to 8 completed tasks with one clearly identified blocker.
i stopped letting my GPT-6 Astra agents retry forever and the system got more reliable one task kept failing because a source page wouldn’t load the agent tried again then again then another route then another browser session from the outside, it looked persistent inside the workflow, one broken page was quietly consuming the entire run so I gave every job a retry budget 3 attempts after that, write the blocker down, preserve everything already finished and move the safe work forward the first time the rule triggered, I expected a worse result instead I came back to 8 completed tasks, 1 blocker and the exact reason the missing task had stopped nothing was hidden nothing spent the night fighting the same page and the rest of the workflow survived that made retries look completely different to me an autonomous agent shouldn’t be measured by how long it can keep trying sometimes autonomy is knowing when a problem is no longer worth owning
This post quotes another post: x.com/lummox_eth/status/2100288151897981255