I don't think they force you to do that. Not trying to be pedantic, but you can also just wait for a single agent to finish its work.

Nowadays I often feel that that is a better spend of my time. Especially because it allows my mind to wonder about the next thing and what I could have missed. Before I was doing the context switching dance I felt I was going at 10x, but realized I was only perhaps going at 2x. Now I have severely reduced the number of threads, mostly to a single thread, and ask smaller questions. My sense is that I am probably now going at 2.5x but with my sanity restored and with a much firmer grasp of what is going on.

I also find myself coding more skeletons and letting an agent finish it, because I am again taking the time to think of the shape of the solution. The end result is better then when I allow the LLM to flail about in the codebase.