The sleep thing is the tell.
You've traded one bottleneck for another. Arguably, the constraint used to be "how fast can I type correct code," now it's "how fast can I validate whether this code is correct."
The first one had natural breaks. The second one doesn't.
Brooks's surgeon model assumed the surgeon's judgment was the scarce resource and everything else was support staff. That's still true, but now the support staff works 24/7 and never stops bringing you things to review. The limiting factor isn't the agent's output. It's your ability to hold the conceptual model in your head while context-switching across ten parallel streams of plausible-looking code.
I've noticed something similar. I'm temptede sessions, burn more tokens. Tempted to keep the pipeline full.
For me, the wins come from the opposite: fewer parallel threads, longer uninterrupted chunks of "is this actually what I want" time.
Maybe it's like Amdahl's Law with a vengeance. The sequential part — taste, judgment, architectural coherence — doesn't compress as well.