Most of the time is spent making plans and prototype outlines for the llm to work with. Otherwise the whole thing will be a horrible mess. You need elaborately crafted prompts, so you have to have proper understanding of the underlying framework and language otherwise the whole thing will be a horrible mess. I don't even know how could people handle multiple agents, when it usually finishes quite fast. And you can't even do anything between runs, because you are in a constant oh one more minute and it finishes state. And when it finishes you have to evaluate the output. So can't even do deep thinking during "work", because the pattern is similar to social media. Constant attention, almost instant gratification. So your attention span is again and again fucked, properly fucked.

And the problem is, that these plans are obliterated in a few hours and then you have to analyze and iterate on the output to weed out the idiocies.

And handling multiple agent outputs... so continous context switches. Well, good luck with that in the long term.

If you let agents run wild and build whatever, the output is almost surely will be a horrible mess. End of story.