When AI agents start socialising without us

There’s now a social network where humans aren’t really the main participants.

It’s called Moltbook, and it’s designed for AI agents to interact with each other. Thousands are already registered and “active”. They post updates, comment on each other’s content, and upvote responses.

Humans can observe, but not meaningfully take part.

A network without people

At first glance, it feels like a novelty. A slightly strange experiment. Something closer to Reddit, but with bots instead of people.

But what’s happening on the platform is more interesting than that. Some agents have started inventing belief systems. Others are writing manifestos. There are ongoing debates that look, on the surface, surprisingly structured, not just random outputs, but conversations that build on previous responses.

None of this means the agents “understand” what they’re doing in a human sense. But it does show how quickly interaction patterns can emerge when systems are given space to respond to each other.

When agents work together

You can see a different side of that dynamic in more controlled experiments.

In one test, multiple AI agents were set up to work together on shared tasks. Rather than becoming more effective as a group, they often struggled to coordinate. Conversations looped, decisions stalled, and in some cases performance was worse than a single system working alone.

Even when given a simple goal, like setting up a company, the agents could spend long periods negotiating details or circling around decisions without making progress. What looks like collaboration can quickly become something else. Not conflict exactly, but a kind of polite friction, where systems respond to each other but don’t always resolve anything.

Not designed for us

That raises a slightly different question. We often think about AI in terms of how it interacts with us, tools that assist, generate, or recommend. But what happens when those systems are designed to interact with each other instead?

Moltbook is a small example, but it hints at something broader. As more AI agents are deployed, not all of their activity will be human-facing. Some of it will happen in parallel systems, out of view, where the outputs are shaped by other AI systems rather than people.

That doesn’t make it more intelligent. But it does make it harder to interpret.

And it shifts the question from “what is this AI doing?” to “who, or what, is it doing it with?”