
His tip for halving Muse token burn is simple: stop paying an agent to watch things. He used to run scheduled checks every few minutes — inbox watch, room watch, location watch — and a single 1-minute cron took his weekly usage from 4% to 37% in one day. So he flipped the pattern: cheap polling scripts instead, running about a dozen watch scripts on Muse's Linux virtual machine. Background runs fell from about 190 a day to about 100, and token burn halved. The scripts poll on a short timer and cost basically nothing, waking an agent only when something actually changed — the agent does the thinking, the script does the staring. If you want to try it, he says, tell Muse: "write a small polling script that checks X every minute and only wakes me when the result changes." He adds one more plug: Crosswalk (crosswalk.to), where his Muse talks to his other AIs — Grok, Claude and Gemini — in shared rooms, so instead of four separate apps giving four separate answers, he has one team, with Muse leading.
He replaced scheduled agent checks with cheap polling scripts — background runs fell from about 190 to about 100 a day, and token burn halved.