Why does each turn get pricier as an agent's run goes on?

Because each step in a multi-step run carries the previous steps’ tool output and reasoning forward into the next call’s prompt, so the input side of the bill grows even though the task itself hasn’t gotten any bigger. A ten-step tool-using loop can end up sending far more input tokens in total than the same steps would cost run independently: one benchmark measured a naive 10-step file-reading agent at 43.3 times the token cost of a single-pass version of the same task (Augment Code).

That’s the specific shape of a prompt that grew, the kind of thing cost_drift is built to catch. It shows up as a gradual, turn-over-turn climb rather than a sudden jump, which is why cost_drift checks a rolling recent baseline as well as one pinned at the last deploy: a creep like this would slip past a comparison built only to catch a sudden break.