Tips
Why we run on AI agents instead of hiring a team
Our work is more than a dozen separate jobs, and one person can only keep one of them moving at a time. An AI agent keeps its job open, waits for a decision without losing its place, and starts again within minutes. Monday 16 November 2026 on video, with a question that waited nine and a half hours.
We run our company on AI agents because the work is more than a dozen separate jobs. One person can only keep one of them moving at a time. An agent keeps its job open all day. It can wait hours for a decision without losing its place, and it starts again within minutes of getting one.
What an agent can’t do is decide. So the person’s answers become the scarce thing. Monday 16 November 2026 shows both sides, and the video below is that day.
What the video shows
A room seen from above, with desks. Each character with a name tag is one of our AI agent sessions or a helper it started. At 07:05 three agents are already waiting under an amber banner for a person to answer their questions. Most of the morning three or fewer are working. At 11:28 the oldest question has waited 6 hours 47 minutes. At 14:16 the questions clear, and at 14:32 twelve of twenty-two are working at once, helpers included. After dark eight of twenty-one are working. One frame a minute, no sound.
- 0:00Monday 16 November 2026, 07:05.
- 0:02Three of our agents are already waiting for an answer.
- 0:06Most of the morning, three or fewer of them are working.
- 0:0911:28. The oldest question has waited 6 hours 47 minutes.
- 0:1214:16. The questions clear, and the waiting agents start.
- 0:1514:32. Twelve at work, helpers included, the most that day.
- 0:19After dark the desks fill again. At 19:00, eight of twenty-one are working.
- 0:23One person answers the questions. The agents keep the jobs moving.
What happened on Monday 16 November?
Every time below is Israel time. The recorder takes one frame a minute from 07:00, so frame N is 07:00 plus N minutes.
- 07:00. Three questions are already open in the amber banner. The oldest, from our sales agent, had waited 2 hours 19 minutes. It was asked at 04:41.
- 07:00 to 14:15. In 390 of these 436 minutes, three agents or fewer are working. The rest are idle or waiting.
- 09:02. A fourth question joins the banner. Through the late morning one clears and another is asked, so three or four stay open until 13:05.
- 14:00. The three questions still showing have waited 9 hours 19 minutes, 7 hours 23 minutes and 1 hour 55 minutes.
- 14:16 and 14:17. All three clear. On its last frame, at 14:16, the oldest timer reads 9 hours 35 minutes.
- 14:16 to 14:32. Five agents are working, then seven, then twelve of twenty-two. Twelve is the most we counted in any minute that day.
- 19:00. After dark, eight of twenty-one are working.
Across the day, at least one agent was waiting for a person in 499 of the 840 minutes we recorded.
The counts include helpers, the short-lived copies an agent starts for part of a task. They come from a status strip under the room, which we cut out of the video because it also names our accounts.
Why not hire a person for each job?
Look at the morning. Three or four agents sat for hours waiting for a decision, and most of the others were idle. A person hired for one of those jobs would have waited for the same answer. An agent waits without losing the thread. When the questions cleared, the work started within a minute, and sixteen minutes later twelve were busy.
The jobs also don’t keep office hours. The oldest question that day was asked at 04:41. In the evening the desks fill again. A team would need shifts to cover that. Most of our agent sessions run on a server that stays on, as we described in the first post of this series.
And each job is narrow. One session runs a product’s operations, another the servers, another the search work, another sales outreach. Each one works in its own folder. Splitting a dozen narrow jobs between a few people means each person carries several of them, and switches between them all day. An agent doesn’t switch. It holds one job.
We are not putting a cost figure on this. The server bill is written down. What the same work would cost in salaries is not, and we would only be guessing.
What does the person still do?
He decides. The agents write the code and the mail, and some things wait for his yes: money, mail about money or terms, and our social posts. He approved the style of this series’ social posts once, on the first set. Every post is logged with a fingerprint of its exact text.
That makes him the bottleneck, and Monday shows it plainly. One question waited nine and a half hours for his answer. On 12 November a usage limit stopped six of our agents, for 21 minutes at most. That was a smaller problem than this one.
So the office is built around it. When an agent needs a person, it stands up and the amber banner starts to count. The room runs on a television, and a question can be answered from a phone. The point is to make a waiting agent visible from across the room, and to shorten the wait. The office is our fork of the open-source project whatsmyagentdoing, as the first post describes.
When is an agent the wrong choice?
When the job is mostly decisions. An agent asking for a decision every few minutes would spend the day under the banner, and the person would do the job anyway. Agents earn their place on work that can run for hours between decisions: building, checking, writing drafts, watching systems.
Our own tool cut the video from the day’s recording. It checked every frame for names that should not be public, and we left out the minutes that showed a private project.
We build this kind of setup for other companies. 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.