I had a nagging worry that I talk to my agents too much, that all my back-and-forth was the expensive part. So I actually counted.
My own inputs across the work came to 446. The machine’s own rows, the tool calls, the system traffic, the echo of agents talking to themselves, came to 53,392. My words weren’t the cost. They were a rounding error.
It changed how I think about working with AI. The waste isn’t in the conversation, the part where the judgment actually happens. It’s in the machinery underneath, and that’s the part worth making lean.
Have you ever measured where your AI spend actually goes, or just assumed it was you?