Today, I can across a rather timely and interesting paper titled Physics of Agents: Statistical Mechanics Predicts Collective Behavior of AI Agents. It looked at what happens when AI agents begin interacting with one another like today’s agentic AI. In this study, the researchers created communities of large language model agents and allowed them to freely interact. Over time, the authors found that the LLM created “collective states that informed the decision-making process.

But there’s more and worth a closer look. On mathematical problems with objectively correct answers, LLM interactions tended to improve “collective” accuracy. In other words, the correct agents exerted greater influence than incorrect ones. And this producing what the authors describe as an emergent truth-seeking tendency.

Now the “opinion” twist: On subjective political questions, conviction could still increase even without an objective truth toward which the group could converge. (Note to self: reread Animal Farm.)

The researchers found that these rather interesting behaviors could be described using ideas from good old-fashioned statistical mechanics that examined the interactions between particles. But in this study, the interactions weren’t atoms in motion or magnetic spins. They were AI agents.

So, down another rabbit hole I went, and found myself interested in what this suggested about intelligence itself. When consequential behavior emerges through interactions among agents, where does that intelligence actually reside?

When Cognition Loses Its Address

Statistical mechanics, scary as it might sound, may help make this clearer. A single molecule doesn’t have a temperature, but temperature appears when molecules move around in a container—the faster the hotter.

In this context, the idea of a “collective state” is particularly interesting to me. When it’s applied to AI agents, it begins to take on a life of its own. Consensus—like temperature— belongs to the collective in a way that can’t be assigned to any one participant.

Of course, collective states aren’t unique to AI. From politics to the marketplace, systems can develop behaviors that spring from those groups. But what makes agentic AI unusual (and fascinating) is that these artificial systems are collectively performing sophisticated intellectual tasks that may be greater than the sum of the parts.

We humans are accustomed to cognition having an address. We locate it somewhere, usually in a person. And even collective human intelligence the thinking subjects remain within the network. So far, so good.

Agentic AI can complicate this and give me reason to pause. With one LLM, we might still consider that model as the source of the answer. But with a network of agents, even locating the process that produced the answer becomes more difficult, if not impossible.

This reflects an idea I wrote about in March 2025 called the Cognitive DAO. I wondered if intelligence might emerge across a decentralized network without residing in any single mind. I described it then as intelligence that could be “expressed, not possessed” and perhaps more specifically, “coordinated, not centralized.” Today, these new agentic systems give that question a bit more credibility and relevance. Here’s the key point: we may be approaching cognition without a defined center.

So, our persistent and pesky question about whether an individual AI can think may be becoming too narrow. Intelligence may increasingly be expressed through its individual architecture and its collective relationships. And maybe the better unit of analysis is beginning to expand from the individual model to the system in which it participates—a technological big bang.

The Human Enters the System

I think it’s fair to say that personal agents will increasingly act and, dare I say, “think” on our behalf. The pathway between a human question and its eventual answer may pass through a network of artificial agents. Importantly, we may not perceive or participate in this new found cognitive dynamic. And if more of the “work” occurs within these networks, our relationship to the cognition producing the answer shifts or even disconnected.

That raises another question around cognitive selfhood, or the sustained presence of the self in its own thinking. Agentic systems make that question more complicated because the boundaries of the cognitive process itself are expanding and becoming harder to define.

Today, AI is beginning to complicate many assumptions about thought and the thinker. If intelligence can increasingly emerge through interactions across a technological system, the question isn’t simply if AI can think. We may also need to ask where the thinking is happening, and what it means for us to remain present within it.