Low is good if you’re working in a tight loop. But more risky for more agentic stuff you want to let cook for 30 minutes or more.
Regardless of the model, running it for hours means that the model will takes decisions and assumptions alone instead of you.
No matter how clever the model is, most problems have multiple valid, invalid and unclear decisions to make, running it for a long time is just picking the first option on everything, which isn't usually what you want
It’s like learning to delegate and let go. The more senior I got the more I had to learn to let other engineers make decision i thought were suboptimal but mostly good enough. That positioned me well to be comfortable with agents. It’s contextual how much I’m willing to give them control and how much to review afterwards.