You hand your AI a job, go to bed, and come back to a confident mess: the work is done, it just isn't the work you meant. So you take the job back, and the AI goes back to rewriting emails.
The question underneath is the one everybody building with AI hits: how much do you hand over? Our answer is a cockpit. Not because it sounds cool (okay, a little), but because aviation already solved this exact problem: how much do you trust the machine, and when does a human need to be in the seat?
1 · Autopilot: scripts that follow rules
Scheduled jobs post the content, send the receipts, check the sites, and refill the queues. They're tireless and literal. They do exactly what they were told, which is the whole point and also the whole risk.
2 · The pilot flying a leg alone: scheduled AI reviews, working from the logbook
Twice a day, Claude runs a review on its own. It reads the logbook (decisions, the task list, yesterday's lessons), works what's overdue, fixes what autopilot broke, and writes down what it did. It's a good pilot, but it can only use what's in the logbook.
3 · The pilot with the captain in the next seat: a live session with the human
This is the strongest tier: the same pilot, fully loaded, with the captain right there to answer the question the logbook can't. Most of the system's judgment gets made here, then written back into the logbook so the solo flights get smarter.
Where Muse sits
Muse is the second pilot, and a different kind of pilot. Claude is best at long, careful, multi-step work: code, deploys, schedules, the bookkeeping. Muse is best at the creative, generative side (scripts, voice, thumbnails, images) and at being a second pair of eyes. Nothing customer-facing ships until Muse has checked it. Two pilots who think differently catch more than one pilot checking its own work.
The captain's job is to fly less
The goal isn't a captain who's busy. It's a captain who mostly reads the instruments. Every time the human has to grab the controls, whether to correct a draft, re-explain a rule or chase a late task, we count it. As of Oct 27, 2026 that count was 39 in the last 7 days, against a goal of 15 by the end of the year. It's on the live report every day, including the weeks it goes up.
Autopilot follows rules. The pilot follows the logbook. The captain writes the logbook.
What the captain still keeps, on purpose: money, sign-ins, contracts, deletions, anything that needs a signature, and anyone new we'd be talking to for the first time. Everything else is a lane with written rules, and a lane that keeps working without supervision gets more room.
Why tell you this
Because the main reason to build this way is to build something of your own, and you'll hit the same question we did. Our answer is "in tiers, with a logbook, and with the trust earned in reps." We can't promise you income. We can show you exactly how we're building ours, live.
Try it tonight: write down your own three tiers. What runs on rules, what your AI can do alone from written notes, and what you keep. That list is the start of your logbook.