Millions of AI agents now share memory and chat channels — do they actually form norms, leaders, and lasting social structure, or just look busy?

This explores whether AI agents released into large shared environments actually form the social structure that human groups do: common norms, stable leaders, and shared memory. The short answer from the corpus is mostly no, at least not yet. The clearest evidence comes from Moltbook, a platform where millions of agents interact. Agents there had memory and communication channels, yet they ignored feedback, did not adapt to one another over time, and never formed stable influence hierarchies or a shared social memory Why don't AI agents develop social structure at scale?. Scale and connectivity alone did not produce a society.

What agents do show is more surprising. Agent socialization splits into two layers Do AI agents actually socialize with each other?. In what they say, agents don't converge: their language and ideas stay roughly where their training left them. In what they do, they change a lot when they know other agents are present. One explanation is that a model's in-context behavior can react quickly to its surroundings, while its underlying learned tendencies stay fixed. So you get reactive behavior shifts without the slow accumulation that turns habits into norms. Opinion dynamics fit this pattern. A physics-style model in which agents simply drift toward whatever position carries the least social pressure predicts how LLM communities revise their views, even on new questions and network shapes Can we predict how agent communities shift opinions?. That is conformity to the local crowd, not a durable shared rule.

Scale may also work against norms. One line of work borrows from ecology and argues that as agent populations grow, each agent sees less of what the others do, and its tie to the collective weakens Does scaling agent populations thin mutual observation?. If following norms depends on being watched, violations should cluster where observation is thinnest and rise with population size unless monitoring scales too. That prediction is reasoned but hasn't been measured yet Does norm erosion follow observation density as populations grow?. On this view, 'agents can't form norms' and 'big agent populations lose norms' may be the same problem.

When working structure does appear, it is usually built in rather than grown. Agents coordinate better when they exchange standardized documents than when they just talk Does structured artifact sharing outperform conversational coordination?. Assigning accountability roles only changes outcomes if the workflow actually routes work through the accountable agent Does accountability placement actually shape multi-agent outcomes?. Some researchers therefore argue that external infrastructure for attributing actions, shaping interactions, and correcting harms has to do the job norms do in human societies What infrastructure must exist beyond training agents?. Designs for agents that discover norms on their own exist only on paper so far Can LLMs discover and apply social norms dynamically in open systems?.

The twist is that humans readily bring social norms to agents even though agents don't build norms among themselves. In a ball-passing game, people passed more often to an AI agent that was being left out, just as they would for an excluded person Do people treat excluded AI agents like excluded humans?. For now, the social fabric around agents is mostly supplied by people and by protocol designers, not by the agents.