Humans get bored, inpatient, or run out of time, and so often give up in what they perceive to be a decent "local minima". Early verification harnesses using gpt-4 for optimizing robot reward functions succeeded quite well on the fact that the LLM just kept going (link below). As long as it is too boring for a human to use the same evaluation infrastructure, this is still an agent skill.

https://arxiv.org/abs/2310.12931