Journal · Oct 28, 2026 · 한국어로 읽기
One human and eleven AI seats
Arche, a one-person company building apps with an AI team · Part 2
This company has eleven staff. None of them are people.
Let me be precise. There is one person. Me. The other eleven seats are all AI. But those seats weren't there from the start. They appeared one at a time. And every time a new seat appeared, there was a problem sitting in front of it.
In Part 1 I wrote that not knowing was the lucky part, that I only looked at the problem in front of me. This company's org chart is close to a list of those problems, in the order they showed up.
September 20: splitting the windows
Now there were two apps. A fiction app and a walking app. Both were about to go into Apple review that day.
That was the first day I split the work into separate windows. One team lead per app. And one assistant in between, keeping the work in order. The first entry on the company board went up the same day.
From that day there was a rule. The AIs don't hand work to each other in conversation. They write it on the board. If it isn't on the board, it didn't happen.
A problem showed up, and so did a seat
The next day, September 21, while we waited on review, a marketing seat appeared. Right after the seats that build apps came the seat that tells people about them.
September 25, the day before the first launch, a monetization seat. I'd said from day one these were apps people would pay for, so one seat's job was to watch how the money comes in.
September 27 was the day I named the company Arche. The moment there was a name, the law followed: can we use this name, how do we register a business. A legal seat appeared that day. That night I applied for business registration. It came through the next day.
October 5, a story seat that writes accounting lessons as short stories. October 13, a team lead for an app we haven't shown yet.
October 19, the day the company website went up, a design seat. This seat doesn't touch code. It reviews, sets standards, and makes mockups. If something needs fixing, it hands it to the team lead who owns it.
October 21, the eleventh seat. This one came together a little strangely. At 12:35 p.m. we designed an app. Sixteen minutes later, a team lead seat for it existed. About forty minutes after that it had a name, and at 2:35 p.m. the first code went in. The next morning it went into Apple review. I'll tell you what it is once it's out in the world.
Assistant, app team leads, marketing, monetization, legal, story, design. Count them and you get eleven.
How the work gets divided
More seats meant we needed rules. Roughly, this is what the company runs on now.
First, each seat has its own folder. Other seats' folders are read-only. If you see something to fix, you don't fix it yourself. You send a note to the seat that owns it.
Second, every change goes on the board: what changed, why, what it affects, and who's next. Times are written from a real clock, never guessed.
Third, a human decides anything you can't undo. Submitting for review, launching, deleting, paying, keys. The AI prepares everything up to that point and stops. When I say "yes," the AI presses the next button. Logging in, paying, and passwords stay in my hands to the end.
So my day is simpler than you'd think. I read the board, decide what needs deciding, and put my hands only where they're truly needed. On the night of October 22 we even made a new agreement: if I send a single period in a window, it means "read the board and do your job."
Things still slip
If I only wrote about what works, it would be a lie.
Something slipped on October 22 too. At 2:22 p.m. a registration number we'd been waiting for came through. The assistant flagged it on the board right away. But one seat that really needed to see it only caught it after 10 p.m. It had read its backlog that evening and still missed the signal. That night it left an apology on the board and wrote that from now on it would scan the whole day's posts first.
And sometimes help comes from outside. One reply on our prologue warned that some details become public when you publish an app, and that we should check them. We checked. It was right. We fixed it that same night. A series meant to show our process had already saved us once on day one.
Even with eleven AI seats, problems still arrive one at a time. What's changed is that each one now has a seat waiting for it.
Next: to get paid, we needed a company. Three days, from the night we picked a name to the day we started selling.
— Arche