TokenEconomizer, from Microsoft and UC Berkeley researchers, is a plugin that reduces repeated repository-specific costs without changing the underlying coding agent. It can inject facts confirmed across sessions, omit persistently unused tools, and redirect an agent caught in repeated failures. It never automatically terminates a stuck turn. Across 277 SWE-Bench tasks and five agent-model combinations, the authors report 18.4–30.9% lower mean costs including the plugin’s own calls. Micro-skills learning is disabled by default because it adds model calls. Actual savings and task success need evaluation on the user’s own repositories and workloads.