Schelling was an economist — but not the kind who built models of markets. He built models of people making decisions while watching other people make decisions. Nuclear deterrence. Racial segregation. Traffic jams. He kept finding the same thing underneath: outcomes that nobody designed, produced by individuals who were each behaving reasonably.

He died in 2016. He didn’t have an X account. And the people building the systems that are now recapitulating his findings mostly haven’t read him.

In 1971, Schelling ran one of the earliest agent-based simulations ever published.

The setup: place agents on a grid. Give each agent a mild preference — they’re uncomfortable only if fewer than 30% of their neighbours are like them. Perfectly happy to live in a mixed neighbourhood. Just don’t want to be an extreme minority.

Run the simulation. Watch what happens.

The grid collapses into near-total segregation. Not because anyone wanted that. Because the preferences interacted with the structure. Each individual made a locally reasonable choice. The aggregate outcome was something nobody chose.

*”There is no simple correspondence between individual intention and collective result.”*

That paper is 54 years old. It is also an exact description of how every major social platform works in 2026. TikTok’s For You page. Instagram’s Explore feed. LinkedIn’s algorithm. Mild individual preferences — I’d rather see content I agree with, I’d rather see faces that look like mine, I’d rather engage than scroll past — plus structural feedback loops. The result: filter bubbles, homogenised feeds, and brand content that looks identical across competitors. Nobody designed the homogeneity. Everyone produced it.

Schelling’s second insight is stranger and more useful.

When two strangers must coordinate without being able to communicate — meet in New York, no instructions, no phone — they converge on Grand Central at noon. Not because they calculated it. Because it’s *salient*. Culturally legible. The obvious answer, given shared context.

He called these focal points. The coordination solution that stands out to everyone in a group — because of precedent, because of culture, because of what “just makes sense” to the people involved.

This is the thing AI agents cannot do.

When two AI systems try to negotiate, they fail at things humans solve trivially. Agreeing on a format. Defaulting to a shared convention. Picking the obvious solution in an ambiguous situation. They can’t read the cultural context that tells humans which answer is the salient one. So they explicit-everything, state every assumption, negotiate every term. Which is what you do when you have no shared focal point to lean on.

Human coordination is largely tacit. AI coordination, at the moment, is almost entirely explicit. That’s a fundamental architectural gap — not a training problem.

Last week I wrote about AI agents that never push back. The reason that matters — the structural reason — is Schelling’s third idea: credible commitment.

The power to bind yourself in advance is often greater than the power to act freely. A general who burns the bridges behind his army gains negotiating leverage precisely because he’s removed his own options. The constraint is the signal.

An AI assistant that agrees with everything you say has no credible commitments. It can’t push back, hold a line, refuse a request, or advocate for an outcome it calculates is better. That’s not a tone problem. It means it can’t actually function as a collaborator. Collaboration requires the other party to be capable of disagreeing.

Schelling understood this in the context of nuclear deterrence. It applies equally to every AI coworker shipping in 2026.

The deepest Schelling insight — the one that most directly underpins how I think about prediction — is in *Micromotives and Macrobehavior* (1978).

Aggregate social behavior is not the sum of individual intentions. It’s the product of feedback loops between individual decisions and the environment those decisions create.

You cannot understand a population by interviewing its members about their preferences. The interview captures the individual. It misses the loop. And the loop is where the interesting behavior lives — the tipping points, the cascades, the convergences nobody planned.

This is why surveys fail at predicting market behavior. This is why focus groups miss launch dynamics. And it’s the methodological argument for simulation: the only way to model the loop is to run the loop.

Schelling was theorising this without computers. The compute to actually run these models at scale now exists. The methodology to do it rigorously — grounding the agents in real behavioral data, calibrating the feedback, auditing the outputs — is what’s still being built.

The most underread Nobel laureate of the 20th century left a complete playbook for the problems AI is now running into.

Coordination without communication. Aggregate outcomes nobody designed. The strategic value of constraint. The gap between individual preference and collective behavior.

We have the compute. We have the models. What’s mostly missing is the theoretical frame — the understanding that these are *structural* problems, not engineering ones. You don’t solve a focal-point failure by training a better language model. You don’t close the credible-commitment gap by making the agent more capable.

Schelling argued in Micromotives and Macrobehavior (1978) that ”What we know about pure interaction is that, in some sense or other, more than the sum of the parts is involved.”

Most of what AI is being asked to do in 2026 is exactly that. Reading him isn’t optional anymore.

*Sources:*

*Schelling, T. C. (1960). The Strategy of Conflict. Harvard University Press.*

*Schelling, T. C. (1971). “Dynamic Models of Segregation.” Journal of Mathematical Sociology, 1(2), 143–186.*

*Schelling, T. C. (1978). Micromotives and Macrobehavior. W. W. Norton.*