We've reduced token costs in Cursor by 7% with no drop in agent quality. Savings came from tighter prompts, selective tool loading, better caching, and compressed file reads. 📷