Coding agents are measured one task at a time and used all day in the same repository, so the same discovery is bought again on every task: profiling 60 tasks found repeated operational actions were 12–28% of tool-call bytes, 16 of the 23 tools on offer were never called once, and the 17% of runs that failed took 32% of the spend. An external controller that pins one-line facts learned from earlier runs, prunes tools nothing calls, and stops runs a judge rules stuck cut token cost by 18.4–30.9% across three agents and two frontier models, with task success unchanged or slightly higher — and cheaper in all 30 repository-and-setting cells. No single one of the three gets you there alone.