A One-Person Company Runs Hundreds of AI Agents, and the Comments Are Better Than the Video

el.cine posted a video titled "this how we run one-person company with 100s ai agents". On screen is a node graph, each dot has a label, and arrows on the connecting lines dart back and forth like a subway dispatch screen. After the video was posted, reposts and comments took off together. After reading the comments, the video itself is just an appetizer.

The first question in the comments is very pointed

Diver Balance asked below: hundreds of agents, so who reviews all that output?

el.cine answered bluntly: different models, different tasks. Someone kept pressing on exactly how the review was done, and he added: another 100 agents on top.

Biscuit asked the same thing: who checks the work they do? The answer is another batch of agents.

Felix Yang took a different angle: is there anything you absolutely would not hand over to an agent? el.cine replied: nothing.

The skeptics were not about to let this video off the hook

Xiaohua said he watched it 108 times and found that several arrows in the diagram were running around aimlessly and not working at all. Daniel Ek's parody account dropped one word: performative. Jonathan Greene asked more sharply: why insist on letting agents do design in a WYSIWYG way? Aren't there better demos? Athulraj T laid out the conclusion directly: isn't this just a Chrome extension called swarm?

Denis B.'s line was the funniest: is this how non-engineers imagine code is written?

The comedy MVP is Clayton Brown: a bunch of agents that do everything sounds great, until one books three dentist appointments and another orders 40 pounds of brisket. Someone still has to be the foreman.

Some people were genuinely moved. Eric Forgy said the video gave him chills, and Dominik Martin said it gave him goosebumps, adding that if this really happens, the era is going to be exciting. el.cine himself said he was very excited. Prayush Kale said he couldn't wait to see a swarm schedule a meeting with another swarm to discuss why nothing got done, and the author replied: how do you know they aren't already in a meeting.

What this video is actually selling

It isn't selling capability; it's selling the first time orchestration has become visible. In the past, agents were a line of log scrolling past in the command line; now it's an understandable dynamic graph, with a label on every dot and arrows showing task flow. This visual shift itself has value: it turns "multi-agent collaboration" from an architecture diagram in a paper into something that can be watched, mocked, and questioned.

But those questions in the comments won't disappear just because the visuals look good.

The core issue is: where did the cost of review go? "Another model reviews it" sounds like a solution, but in reality it just moves the judgment into another black box. If 100 agents do the work and 100 agents do quality control, how do you calculate the marginal cost of this model? The video has no answer. What's more troublesome is that the probability of review errors compounds; when models review models, errors don't cancel each other out—they only endorse each other.

The second issue is the 40-pound brisket problem. Agents can do ten thousand things in parallel, but the constraints of the real world are serial. The dentist has only one appointment slot, and only one fridge can hold the brisket. The foreman role can't be swarmed away; it will only become more important.

A one-person company is not a zero-person company

Every time this kind of demo appears, it gets read as "people are no longer needed." What it actually says is one thing: the execution layer can be spread very thin, while the decision layer becomes more concentrated. el.cine said there is nothing he wouldn't hand over to an agent, and that itself is telling, because whether to hand it over is his judgment to make.

The truly scarce thing has never been output; it's judgment. To get a company running, what you need isn't hundreds of hands doing work, but a person who knows what should be done and when it counts as finished. No matter how big the swarm is, someone still has to stand nearby and ask whether the job is actually done.

Some of those arrows in the video are running around aimlessly, and that's actually quite real. In a real multi-agent system, some part is always spinning idle, confirming with each other, waiting for an input that will never arrive. A demo won't show you this; the comments will.