Let me start with a conjecture: the way we work is about to change. Instead of carrying a computer around, you leave the computer at home and take only a screen. Inside that screen lives an AI steward. You talk to it, and it runs the real computer back home for you. This is not speculation. I have just run the whole loop, end to end.
The pattern has three layers.
At the top is me, with a screen. I do exactly three things: set the rules, make the calls, and accept or reject the result.
In the middle is Muse, the AI steward that travels with me. It knows my context. It breaks work into tasks, writes the instructions, watches the progress, and checks the quality.
At the bottom is the computer at home, always on, connected to the outside through Tailscale. No shared local network, no machine exposed to the public internet, and no poking at a remote desktop one click at a time. Muse can step into that machine whenever it needs to: see what it is doing, give it orders, and bring the finished work back. On that machine sits an execution AI, Antigravity, which does the actual hands-on work.
Me (the decisions) → Muse (the management) → the computer at home (the execution). The whole route is that simple.
The evidence is a 37-slide proposal deck finished in a single day. I was out the entire time, and only tapped "approved" on my phone a few times:
- The execution side burned through its credits at midnight and stalled. Muse worked out the billing rules and brought the decision to me. I approved a 25-dollar prepaid usage budget, and it found the switch buried in a command-line settings file and turned the work back on. Total actually burned: under 8 dollars.
- During acceptance, checking slide by slide against the source text, it caught a chemical formula the execution side had fabricated. The same error was hiding on three slides. It was cleaned out before delivery.
- The finished deck had too few figures. It traced the cause: the reference I had given was full-page screenshots, and you cannot peel an image off a screenshot and reuse it as material. So it extracted all 296 original images from the template and every figure from the proposal document, built a proper asset library, and ran a dedicated round to put them back.
- The deck came back to my phone. I read it, nodded, done.
That whole day, I never touched that computer's keyboard. It was at home. I was out. Between us stood one AI that knows how to manage things.
But the deck also hid a serious pitfall, one that nearly wasted the entire day. When the finished file was opened in real PowerPoint, it popped up a "needs repair" warning. After the repair, only the first 3 of the 37 slides remained; the other 34 had been hollowed out completely. The strange part: everywhere else, the file looked perfect. The preview images were normal. It exported to PDF normally. Another app opened it normally. Even Microsoft's own validation tool reported zero errors.
Muse dug for a whole night and finally located two invisible faults. First, the "packaging" was off-spec: the file had not been packed by PowerPoint itself, and it lacked a set of format markers PowerPoint expects from its own files, so PowerPoint refused to recognize it. Second, a misrouted wire was buried inside: from slide 4 onward, every slide carried an inconspicuous internal reference pointing at the wrong target — at a layout file. The execution tool had cloned parts from the template and carried the reference numbers over unchanged, but in the new file those same numbers already meant something else. The first three slides had correct references, which is why only they survived. Once both faults were fixed, I double-clicked the file myself. It opened cleanly, first try.
It taught me something: with AI-delivered work, there is an invisible layer between "looks finished" and "actually works." Automated checks can pass everything. Only opening the result in the real software, by hand, counts as acceptance. That is also why the top layer of this pattern — the human accepts the work — is not decoration.
So I have come to believe this: the computer is turning into a home appliance — a box that stays on, holding compute and files. What travels with you is the screen, and the AI inside it. What you carry is not computing power. It is a steward who knows you and can mobilize everything at home on your behalf. And the human role changes too: no longer the person who operates the software, but the person who sets the rules and makes the judgments.
It is still rough. Credits burn out. Files hide faults you only find at the byte level. Some work still needs a human watching the acceptance. But the route is open. The rest is engineering.
← Back to Blog