Running a company with AI agents: one working day on video

A company run by AI agents looks like a small office where one person makes the decisions. Ours has twelve agent sessions. Each one owns a single job: a product, the servers, the search work, the mail.

We recorded Thursday 12 November 2026 from 07:00, one frame a minute. The video below is that day up to 19:00. Things went wrong in it, and they are still in. One agent’s question waited more than four hours. Around noon a usage limit stopped six of the agents, and for twelve minutes nobody was working.

Watch this video on its own page

Who is in the room?

Each character with a name tag is an AI agent, and the tag is the name of its session. A session is one long conversation with a coding agent, working in its own folder on one job. Ours run on Claude Code. Twelve long-running sessions were on screen that day.

Session	What it looks after
coo	Operations lead. It keeps the other sessions running and passes a new rule to all of them
synapse	Engineering. This website, its search work and our company mail
tennis ops	Operations for Live Tennis API, our tennis data product
fax for agents	Product work on Fax for Agents, a fax service that AI agents can use
the accounting session	Quality for an accounting product
seo mania	SEO. The tool that plans our daily search and link work
feedback watch	Support on a client project
supervisor migration	Infrastructure. A move to a new server
fleet sweep	Maintenance
mrtrader desk	Trading research
mrshort	Trading research
mrfiverr	Sales outreach

Nine of them run on a rented server with 8 dedicated virtual CPUs and 32 GB of memory. It stays on, so the work continues when the laptop is closed. Each session there is a service that restarts itself. We stopped one by force, and 6 seconds later a new process had picked up the same conversation. The other three run on the laptop.

The extra characters that share a tag are helpers. An agent with a large task starts short-lived copies of itself for parts of it, and each copy gets a desk near its parent. Some helpers are a second coding agent, Codex, and those are drawn as robots.

Some figures are not agents. The ones on the bench by the door stand for customer matters that are waiting for a decision. The small figures outside the windows are visitors on our sites.

What happened that day?

Every time below is Israel time. Frame N of the recording is 07:00 plus N minutes.

• 07:00. The recording starts. Four of the seventeen characters are working, helpers included.
• 07:02. A red pill appears. It says the automated checks of Fax for Agents are failing, and it comes and goes for the rest of the day.
• 07:49. The session called seo mania asks a question it can’t settle by itself. Its character stands up, and an amber banner starts a timer.
• 11:54. Two agents reach a usage limit on their plan. The AI service allows a set amount of use in a period, and they have used it. Four more follow.
• 12:02 to 12:13. Nobody is working. The stopped agents stand at the water cooler with empty cups.
• 12:15. The banner clears and the agents start again. Six of eleven are working. On its last frame, the question’s timer read 4 hours 25 minutes.
• 13:31 to 17:40. Three new versions of Fax for Agents begin their release, and a new version of our SEO tool goes out.
• 15:00 to 18:00. A quiet afternoon. In 129 of these 181 minutes, one agent or none is working.
• 17:44. Fireworks in three windows. The office lights them when a product gets a new version, and Fax for Agents 1.17.2 had been tagged at 17:40.
• 18:40. Ten of twenty-one are working. The hour after 18:00 is the busiest so far.

The limit stop lasted 21 minutes. The question took far longer. For over four hours one job stood still, waiting for a person.

The counts come from a status strip under the room. We cropped it out of the video, because it also names our accounts.

What does the person do?

One person works here: Ben, who founded the company. The agents write the code and the mail. He decides, and some things wait for his yes.

• Money. A refund or a change to what a customer pays is drafted by an agent and decided by him.
• Mail about money or terms. Routine replies to customers go out by themselves. A reply that touches money or terms waits as a draft until he approves it with one letter.
• Social posts. Each one is approved first, and the approval is logged with a fingerprint of the exact text.

That is why the amber banner matters. An agent that needs a decision cannot go on without it. So the most useful thing the screen does is show, from across the room, that someone is waiting.

Why is it drawn as an office?

Twelve terminal windows can’t be read at a glance. A room can. You see who is working, who has finished and who is waiting for you. It runs on a television in the room.

The office is our fork of whatsmyagentdoing, an open-source project under the MIT licence. Its 3D models are by Kenney, released as CC0. On 11 and 12 November we added what a company needs on top of it. Sessions from two machines share one room. The banner has its timer. We built a page for answering a question from a phone. And the day is filmed.

Nobody edits the daily film. A recorder takes one picture a minute between 07:00 and 21:00. At 21:10 a timer turns the day’s pictures into a 30-second video. No model picks the moments.

For this post we changed it by hand. We stopped at 19:00, because a client’s project came on screen after that. We dropped three frames that showed an alert about one of our sites. We slowed the opening and the minutes around noon, and moved in close for three shots. And we cropped the status strip.

What comes next?

This is the first note in a series on how the setup works. The next ones cover the rules each agent follows, the checks before anything ships and what broke along the way.

We build this kind of system for other companies too. One service covers agents that act in your tools, and another covers the internal workflows around them. If you have a job in mind, tell us about it.