I try to be creative with them. I sometimes ask the AI questions like how to take the thing I am working on to the next level, "ultrathink", "see the bigger picture", and those runs usually give better results than the narrow ones, and they cost tokens. Most of mine go to reading, not writing: a plan review last week spun up about 26 agents reading the repo and git history, then a skeptic agent tried to refute every answer with file:line and overturned seven. One runaway run hit around 76 agents and ate a whole 5-hour window. That is what made me cap parallelism and be more observant of my usage.