Building in public
I build this company alone.
So I gave it five
AI colleagues.
Not five assistants that agree with me. Five that each know different things, own different parts of the company, and argue with each other and with me. This page explains how that works, shows a real argument between them, and is honest about where it fails.
- 5
- AI colleagues
- 1
- human, who decides
- 0
- things they do alone
Start here
Five questions, answered plainly, so the rest of the page makes sense.
What is the company?
AroundMe puts small groups of four to six people together for a real meetup: a set place, a set time, people who want to do the same thing. It launches at a university in California this autumn.
Who works on it?
Me, in the evenings. There are no employees. That is the constraint everything below is a response to.
So what is an “AI colleague”?
Instead of one assistant that answers everything, there are five. Each has a job title, one area it owns, and its own set of documents it is allowed to read. They are separate: none of them sees what the others are thinking.
Why give them different information?
Because an assistant that knows exactly what you know will agree with you, and agreement is worthless. Give five of them different pieces of the picture and they reach different conclusions, then argue. The arguments are the useful output.
What can they actually do?
Read, analyse, draft, and recommend. They cannot publish, send, spend, or decide anything. Every real decision is still mine, and that is deliberate.
The five of them
Each one owns exactly one area. If a job belongs to someone else, they say so and hand it over rather than quietly doing it themselves. The names are a private joke; the job titles are not.
Gilfoyle
CTOThe code and the systems that run the product.
Tells me what an idea costs before I get attached to it.
Jared
COORunning the actual meetups and the university launch.
Refuses to ship a plan without a backup for every step.
Neha
CMOGetting people to hear about us: writing, search, campus.
Argues loudly about what waiting is costing us.
Donna
Chief of StaffThe numbers, and what they honestly do not prove yet.
Asks “compared to what?” before agreeing anything worked.
Tom
Trust & SafetyAnything touching user data, privacy, or safety.
Decides nothing, by design. His tools will not let him.
Who is allowed to read what
This is the part that makes the whole thing work, and it is the part people skip. If all five read the same documents, all five give me the same answer. So most documents are deliberately given to only one or two of them. A filled circle means that colleague can read that document. An empty one means they cannot.
| Document | Gilfoyle | Jared | Neha | Donna | Tom |
|---|---|---|---|---|---|
| What the product iseveryone | |||||
| The task listeveryone | |||||
| The decision logeveryone | |||||
| The code plan | |||||
| The launch plan | |||||
| What we promise users | |||||
| How matching works | |||||
| The article plan | |||||
| Website traffic data | |||||
| Safety rules |
The three shaded rows are the honest catch. Everyone reads what the product is, the task list, and the decision log, because a company where nobody shares a source of truth is not a company. But it means when all five agree on something from those three, that is one opinion repeated five times. I count it as one.
A real meeting
Watch them argue
The question on the table
We have about 90 finished articles sitting unpublished, written to bring people to the site from Google. Publishing them costs nothing. Should we publish them now, or wait until the real cafes and venues for our first meetups are confirmed? I asked all five.
Publish them now. An article earns nothing until it is live, and ours have been sitting finished for weeks.
Worth knowing before you decide: the system posts one article a day by itself, and only four are lined up. In four days it goes quiet whatever we choose.
Careful. Some of those articles name specific cafes as places to meet. We cannot publish those until the cafes have actually agreed. The rest are fine.
Publish, but do not expect it to bring people in. The whole site has produced six signups so far. Nothing in the data says article number 61 changes that.
Some articles describe a safety feature we have not built yet. Those need checking before anything goes out.
You both said publishing had stopped working. I checked the server. It had not. Please verify before asserting.
Your count of 88 articles is wrong. Twenty-eight of them name real venues, so they are blocked too. The number is about 60.
You said only seven are blocked. More than that promise users a reporting button that does not exist yet. We cannot publish a safety promise we are not keeping.
You called this free. It is not. Someone has to read 60 articles before they go out, in the same weeks we are signing venues.
Conceded, all four. A number with a source behind it beats a number I remembered. Amending my recommendation.
Agreed: publish the articles that do not name a venue, hold every one that does until the venues are real, and check the safety claims first. About 60 articles, not 88. My count was wrong.
For the human: four yes-or-no questions, in order of urgency. Nothing was decided without me.
Four of the five disagreements that round were aimed at the one writing the summary, and it conceded all four. A meeting where the most senior voice is never wrong is not a meeting.
What it has actually done
Two things it caught that I would have missed, and one it got wrong.
An article about to publish a false claim about our own code
One article said our software caps groups at seven people. It caps at six. The article was days away from publishing itself automatically, and the checklist attached to it for fact-checking repeated the same wrong number. A document that checks itself will approve its own mistakes.
Two of them designing against each other
Part of our launch is a five-minute demo run live in a university lecture hall: students scan a code and are sorted into groups on the spot. Tom rejected one idea for it as a privacy risk, correctly, and suggested a replacement. His replacement added an AI step into a moment that has to finish in under three seconds in front of a professor, which Gilfoyle had already ruled out. Neither could see the other's work. Nothing was catching that, which is why the meeting above now exists.
Giving one of them too little
Donna was given only the numbers section of the launch plan, not the schedule that goes with it. She concluded, with total confidence, that we were running late and should freeze everything else. Jared, who had the whole document, showed we were not late at all. Half a document produced a wrong answer stated far more confidently than no document would have.
Where it falls short
The version of this story where five AI executives run a startup by themselves is a better story, and it is not true. What is true:
It is one AI wearing five hats.
All five run on the same underlying model. Giving them different information makes them notice different things. It does not make them five independent thinkers, and anyone telling you otherwise is selling something.
They share more than I intended.
Three documents are on all five reading lists. So when all five agree about something from those, that is one opinion repeated five times, not five sources agreeing. I count it as one.
The first version never actually argued.
In the first real test, not one of them challenged another. Every disagreement was politely handed to me to settle instead. Agreement that costs nothing proves nothing, which is exactly why round two now forces each one to name something it thinks another got wrong.
They forget everything between sessions.
Nothing carries over. All continuity lives in written files, which is the real reason the decision log matters more than any of the clever parts.
The useful lesson is not that AI can do tasks. It is that getting anything out of it is an organisational question: what each one is allowed to see, what it owns, and when it is made to argue.
Why a friendship app has an AI org chart
AroundMe launches at UC Riverside this autumn. It puts small groups of verified people together around something they actually want to do, at a real place, at a set time. That is the least automatable thing there is: cafes have to be visited, professors have to be asked, and someone has to stand up in a lecture hall. The five of them handle everything that happens on a screen, so my hours go to everything that happens in a room.
How AroundMe works · How the matching works · Join the waitlist