Audit my recent agent sessions. Where did the tokens go? Split into: real work, re-reading context, repeated checks/builds, reports. Name each redundant loop and how to cut it. Then have an independent subagent check your findings for anything you missed.