OpenAI documents what breaks when a model runs unsupervised for hours
A model executing a long task accumulates invisible errors that wouldn't show up in a short run: goal drift, unplanned actions, pointless loops. If you're delegating long-running tasks to an agent, this write-up tells you what to watch for before you let it run unattended.
practitioners › Iterative rollout at OpenAI, guardrails adjusted after incidents observed in production, no publication date specified.
OpenAI →