The cost of micromanaging an AI
The expanding brain usually runs the other way: the further down you go, the more enlightened you are supposed to be. Here it just tracks the invoice.
What prompted this was watching how fast the tokens add up once you start going back and forth. Obvious when you say it out loud, less obvious while you are doing it. An agent left alone works through one long, warm context: it reads the files once, holds the plan, keeps going. Every time you cut in, that run ends and a new one starts. It re-reads, re-orients, re-explains itself to you, then picks the thread back up. “Is it done yet?” is not a free question. It is a full turn, and it drags the whole conversation along behind it.
Pair programming is the expensive version of the same thing. You are in the loop on every step, so every step pays the toll.
Which is the honest argument for a Max plan, or whatever the top tier is called on the tool you use. If you are in there every day, the back and forth is not an edge case, it is how you work, and a small package disappears in an afternoon. A flat fee once a month beats rationing your own questions.
None of this means walking away and trusting the output. The check at the end is worth whatever it costs. But reviewing work and hovering over it are different things. One happens once. The other happens every ninety seconds, and it shows up on the bill.