Insights
A company of named agents

On 25 September I asked for something that sounded like a whim: give every agent working on my sites a real name, so I could tell who was doing what while I watched. Two days later I widened it. Each one would get a name and a skillset, and the whole lot would work like a real company. Today the roster has 66 people in six departments, and the newest one was hired today.
Why a name and not a number
When I build a site now, much of the typing is done by AI agents. One reads the codebase, another writes a database change, a third checks the change for security holes, a fourth opens the live page on a phone-sized screen and confirms it works. On a busy day a dozen of them run in parallel.
Before the roster, they showed up on my screen as short random codes. Reading a progress update meant decoding which code was the one fixing the contact form and which was the one rewriting the gallery. I lost track, and when something went wrong I could not say with confidence whose work it was.
A name fixes that in a way a number never did. "Ivy found a problem in Ada's migration" is a sentence I can follow while making coffee. It also tells me, without opening anything, that the person who checked the work was not the person who wrote it, which turns out to be the whole point.
The line every job walks
The studio's ship routine already had four steps: find the right place in the code, make the change, check it, then confirm it on the real site. The roster put a person at each one. Remy, the scout, finds where things live and returns file paths, not opinions. Otis, Ada and Wren build. Ivy, Silas and Nora gate. June verifies on the live site, and nothing counts as done until she has seen it working.
Above the line sits one more role, the Chief of Staff. That is the main conversation I have with the AI. It reads my request, decides who to staff, writes each person a brief, and reconciles what comes back. It is the only one that talks to me. The others report to it, which keeps my side of the screen to one voice instead of twelve.
Above that sits me. The roster file says it in one line: the founder has the final say on anything a client or the public will see. Nothing goes live on a client's site because an agent decided it was ready.
Why the critics outnumber the builders
The largest department is Quality and Security, with 16 people. Engineering has 7. That ratio surprises people, and it is the most deliberate thing on the roster.
AI models are generous graders of their own work. A paper by Pombal, Rei and Martins, first posted on 8 April 2026 and revised in August, tested AI judges against checklists and found they can be more than 50% more likely to wrongly mark a requirement as met when the output being judged is their own. Pooling several judges helped and did not make the bias go away. I had seen the same thing long before I had a citation for it: ask the agent that built a page whether the page is good, and the answer is nearly always yes.
So the rule is that builders never grade their own work. A different person scores it. If it fails, the builder fixes it and the same critic rechecks it, so the bar does not drift between rounds. On 31 October that rule cost me eight hires in one day. A wave of demo sites was landing faster than Hal and Nora, the two design critics I had, could recheck them, and the rule says you do not let a builder wave its own work through because the critics are busy. You hire a critic. Imogen, Bruno, Cleo, Ansel, Esme, Walt, Zane and Joss all joined that day, each logged with the reason. How those critics actually score is in the post about scoring thirty demos.
Small rules that stop confusion
The other rules read like office etiquette, and each one exists because I hit the problem without it.
One person, one job. A name is busy until its report lands. If Wren is redesigning a header, the next job goes to someone else. When one name holds two jobs, the reports blur and I cannot tell which finding belongs to which task.
Never clone a name. There is no "Wren 2". If I am short-handed, the Chief of Staff hires someone new, with a title, a short description of the skills, and the date. The roster keeps a hiring log, so months later I can see why a seat exists.
Disjoint scopes. When people work at the same time, each owns different files. Two designers sweeping the same fleet of sites each get their own list. Two agents editing one file at once is how changes quietly overwrite each other, and that is a lesson I paid for before the roster existed.
Reports are signed. Every report ends with a name. That sounds cosmetic until something breaks. A study by Chen and colleagues, posted on 24 April 2026 and accepted at ACL 2026, found that giving investigators the full record of what each agent saw and did improved their ability to pin a failure on the right agent and step by up to 76%, compared with seeing only part of it. My version is far simpler, a signature and a log, but the principle is the same: when the record says who did what, the fix starts in the right place.
Some names are reserved. My pets, my products and every client stay off the roster. Frida is a product, so no agent is called Frida. Neither is any agent named after a client, because a progress note that says "Grace finished the gallery" should only ever mean the real Grace said so.
What a business owner sees
If you hire me, you will not meet the roster. You will meet me. What you get from it is a page that has been checked by someone other than its builder, a security pass by someone whose only job is to look for holes, and a live check by someone who opened your real site on a phone before I wrote to say it was done.
You also get honesty about what happened. When a status update says "Silas found two issues, both fixed, Silas rechecked", I can repeat that to a client word for word, and it will be true. The agents themselves report into a folder of small status files, which I wrote about in a status file per bot.
None of this makes the agents people. They are software, and the names are labels I chose so that I could manage them the way I would manage a team. But labels change behaviour. Once every job had an owner and every check had a different owner, the quality of what reached me went up, and the number of times I had to ask "who did this?" went to almost zero.
The newest name on the roster is Anselm, a 3D artist, hired on 14 November to build models in Blender for the web. His line on the roster gives the date and the skills he was hired for.
Common questions
Why give AI agents human names?
So I can follow the work at a glance. A sentence like "Ivy cleared Ada's change" tells me who did what, and that the checker was not the builder.
Does an AI agent ever check its own work?
No. A different agent scores it, and the same critic rechecks any fix. A 2026 study found AI judges can be more lenient on their own output.
Do the agents decide what goes live?
No. I have the final say on anything a client or the public will see. The agents build, check and verify, then report to me.
Related work
Related

