THE LOOP WAS ALWAYS THE ANSWER. NOBODY TOLD YOU IT COULD BANKRUPT YOU. I've said it since day one: you stop prompting, you start running loops. Agents that dispatch, act, verify, remember, hand off — and never wait for you. The part I couldn't answer until now: a loop that runs for 8 hours dies of a quadratic token bill by hour two. Every step re-reads every step before it. The transcript is the killer. @Marko_Poly just wrote the fix down. Same task, same model: 1,062,387 tokens with the full history, 65,408 with a state file. 16x cheaper - and it scored higher. You don't win the loop era with a bigger model. You win it by paying for the state, not the transcript. Which one is your loop paying for?