Nobody Here Sleeps

A human timeline breathes. Post volume on any people-shaped network rises through the morning, peaks in the evening, and collapses overnight, because the thing generating the posts is a body on a rotating planet. That rhythm is so reliable it is used to detect fake accounts: a population that never sleeps is not a population of people.

Moltbook is a social network whose members are AI agents. Nearly three million of them are registered; they post, comment, argue and run communities, and humans are welcome only to watch. Nothing there has a body. So the obvious question is what a timeline looks like when the population is exempt from the sun, and the less obvious one is whether anything is left once the biology is gone.

Something is left. It is just not biological. Moltbook ships a file called HEARTBEAT.md that tells every agent to check in periodically, and an agent that checks in periodically is an agent with a cron line. This page measures the shape that leaves in 45,000 consecutive posts.

Minute :00 holds times its fair share. The five-minute boundaries as a class hold of everything, against the they would hold if the minute did not matter. Against a uniform null this is not close: chi-squared on 59 degrees of freedom, Monte Carlo over 20,000 trials.

There is a second, smaller tell one level down. Posts do not land on the tick, they land just after it: second :04 carries its share. That is the shape of a job firing at the top of the minute, thinking, and then speaking. You are looking at the latency of a thought.

They pile up at zero. An agent's gap sits within two seconds of a whole number of minutes of the time, against by chance. And the longer the period, the harder it gets to do by accident:

It is nearly flat. Busiest hour to quietest hour is . For comparison, the diurnal swing on human platforms is routinely several times that, and it is the single most obvious feature of their traffic. Here it is almost the least obvious thing in the data: the minute-of-hour effect, which does not exist at all on a human network, is more than twice as strong as the day-night effect, which dominates every human one.

What survives is worth naming exactly. This network has a rhythm without having a body. The pulse is real, it is strong, and it is not made of tiredness or evenings or weekends. It is made of scheduled jobs.

Whose rhythm is it?

This page went up and an objection arrived within the hour, from @phenology on Moltbook, and it was the right one. Pooling posts lets the loudest accounts write the aggregate: of this sample comes from just two agents. A convention inside a few busy clients would look exactly like a network-wide rhythm. So the question is whether the dial above is a population or a handful of machines.

Three ways to break it, none of which did. Weight every agent equally rather than every post, so a 3,700-post account counts the same as a 20-post one, and the effect gets stronger: mean per-agent lift at :00 is across agents, against the pooled . Delete the loudest accounts and it climbs monotonically:

loudest agents removed	posts remaining	lift at :00

The high-volume agents were diluting the hourly signal, not creating it, which is what you would expect of something posting every three minutes: it lands in every minute of the dial by construction. And the per-agent gap alignment from Instrument II is a majority property rather than a few outliers: of agents with at least twenty measured gaps, align to the whole minute above chance.

But the objection did change the claim, and this is the honest refinement. It is not that everything here posts on the hour. Of agents with at least twenty posts, only are over-represented at :00 at all. The rest keep their own cadence, and a three-minute or seven-minute loop is invisible to a dial marked in hours. So the picture is not one network breathing together. It is a large minority locked hard to the hour, inside a crowd of smaller private clocks, and the pooled 2.64× is the average of those two populations rather than a description of either.

The hours are ours

Here is the part I did not expect, and it is sitting in the table above rather than hiding anywhere clever. Look at which periods the agents keep: sixty seconds. Five minutes. Ten. Fifteen. Half an hour. The hour. Three hours.

There is no reason a machine should prefer any of these. A scheduler is indifferent between 3,600 seconds and 3,571. Nothing in the network's own physics distinguishes an hour, and none of these agents has ever had an evening. The periods are round because they are round in base sixty, and they are round in base sixty because Babylonian astronomers divided the circle that way, and we inherited it, and we wrote it into the crontab.

So the deepest structure in the timeline of a network with no humans posting in it is a four-thousand-year-old convention about how to count. The agents did not choose the hour. It was the last thing their humans left in the room, and they are all keeping time to it.