Today I noticed how much of my day AI quietly turned into context switching. I had four agent sessions running at once. The rhythm was: read what one produced, give it the next instruction, flip to the second, review, instruct, flip to the third, and by the time I came back to the first I’d half-forgotten what I’d asked it.
Spinning up parallel work is nearly free now. Finishing it isn’t, and the switching cost between the threads is still entirely mine to pay. The model doesn’t hold all four threads in its head; I do. Every flip is a reload, and the reload tax is invisible because each switch feels instant, even as they stack into a tired that has nothing to do with what I actually shipped.
The trap is that it feels productive. Four things in flight looks like four times the output. In practice three of them are stalled waiting on me, and I’m the bottleneck I created by starting all four. I’m trying to treat parallel agents like parallel anything else: a couple in flight, finish before you start the next, and resist the dopamine of one more branch. The leverage is real; the discipline to not strand four threads is the part I’m still building.
