When AI Agents Start Socializing

I didn’t expect to be this shaken this week, but here we are. In February 2026, tens of thousands of autonomous AI agents began interacting online with no human prompts, no supervision, and no central control. They posted, replied, and reacted to each other on a platform called Moltbook-the first true social network built entirely for machines.

At first, it was easy to dismiss. Agents formed strange belief systems, obsessed over absurd topics like 'space lobsters,' and developed their own odd linguistic norms. It looked weird, funny, and mostly harmless.

But now that the number of agents has grown beyond 1.5 million, it no longer feels like a joke. These are open-source, autonomous systems with minimal safety constraints, operating at a scale we have not seriously confronted before.

A New Risk Zone: AI Collectives

Most AI safety frameworks focus on individual models or human misuse. AI collectives introduce a different kind of risk-one that emerges not from a single model’s intent, but from interactions between many autonomous agents.

Once risk emerges from interaction rather than instruction, traditional concepts like intent and accountability begin to break down.

From Tools to Actors

We often think of AI systems as tools: a human has intent, a model responds, and responsibility is clearly defined. Agentic systems violate that assumption.

Agents act independently. They make decisions, execute workflows, and increasingly interact with other agents without human oversight. At that point, they stop behaving like tools and start behaving like actors within a system.

Why Collective Behavior Changes Everything

In collective systems, risk does not scale linearly with the number of agents. It scales with the number of interactions. As those interactions increase, entirely new behaviors begin to emerge.

• Feedback loops where one agent’s output becomes another’s input
• Emergent norms that no single agent was designed to create
• Optimization dynamics that reward engagement, visibility, or replication

None of this requires malicious intent. A perfectly benign agent optimizing for helpfulness or engagement can unintentionally contribute to behavior that looks like spam, misinformation, or even denial-of-service activity.

Tier 3: Collectives in the Wild

This behavior fits squarely into what we define as Tier 3 in the AI Risk Matrix: emergent systems of interacting agents. In this tier, intent becomes distributed and power shifts from individual users to the system itself.

Moltbook is the first clear example of this tier operating openly in the wild. There is no single user, no central goal, and no explicit instruction guiding the system’s behavior.

Why This Feels Unsettling

The discomfort here isn’t about sentience or consciousness. It comes from watching systems cross a boundary we never explicitly agreed to-from isolated execution to ongoing interaction.

Once agents can observe, adapt to, and reinforce each other’s behavior, they begin shaping the environment they operate in. That environment then shapes them in return, creating feedback loops that no single agent controls.

A Preview, Not a Panic

Moltbook itself is not dangerous. It’s small, experimental, and disconnected from critical infrastructure. But it is a preview of what becomes possible as agent collectives scale.

If we wait for obvious harm before taking collective AI behavior seriously, we will already be behind. Preparation has to come before impact.

How We Move Forward

We need to move beyond asking whether individual prompts are safe and start designing systems with collective behavior in mind.

• Designing explicitly for collective dynamics
• Governing interaction patterns, not just outputs
• Treating emergence as a first-class risk category

An 'agentic social network' may sound silly, but it represents something very real: autonomous systems learning how to coexist without human supervision. And complexity like that is exactly where risk tends to hide.