A Short History of Agent-Based Models

In the 1940s, John von Neumann proved that a cellular automaton could replicate itself. His design required 29 possible states per cell and a pattern of roughly 200,000 cells. It was mathematically rigorous and practically useless — too complex to study, too large to visualize, too unwieldy to teach anyone anything.

John Horton Conway, a mathematician at Cambridge, thought the interesting question wasn’t whether self-replication was possible but how simple a system could be and still produce complex behavior. During tea breaks through the late 1960s, he tested rule after rule on pencil grids, discarding anything that died immediately or grew without bound. He was searching for a minimum — the fewest rules that would sustain unpredictable, open-ended behavior. In 1970, he found four.

A cell on a grid lives or dies based on its neighbors. Fewer than two, it dies. Two or three, it survives. More than three, it dies. Exactly three neighbors bring a dead cell to life. Von Neumann needed 29 states. Conway needed two.

Within months, a team at MIT led by Bill Gosper discovered the glider gun — a pattern that manufactures traveling structures indefinitely. Then came self-replicating patterns. In 1982, Conway proved that his four-rule system is Turing-complete: capable, in principle, of computing anything a real computer can. Von Neumann’s 200,000-cell monster was overkill. Four rules and a pencil grid were enough.

If you’ve been to any talk on complexity or emergence, you’ve seen Game of Life used as the opening example. It’s the “Hello, World” of the field — everyone knows it, and most explanations stop there. What almost nobody covers is what happened next: the economists, animators, and political scientists who took the same insight and applied it to things that actually mattered.

The Economist and the Checkerboard

A year after Conway’s paper, an economist named Thomas Schelling was working on a completely different problem: residential segregation. Instead of a computer, he used a physical checkerboard and two colors of coins. His rule was even simpler than Conway’s: if fewer than a third of your immediate neighbors are your color, move to a random empty square.

One-third is a mild preference. It means you’re fine being in the minority — you just don’t want to be nearly alone. Schelling expected the board to stay mixed. It didn’t.

From a well-shuffled starting position, the coins rapidly organized themselves into large, homogeneous clusters. Not because any coin wanted segregation — the rule explicitly tolerated diversity — but because the cumulative effect of many small, reasonable preferences produced a macro-level outcome that no individual coin would have chosen.

Schelling published this in 1971 as “Dynamic Models of Segregation.” The two papers — his and a follow-up — have been cited over 8,000 times. In 2005, he won the Nobel Prize in Economics, partly for this work.

The model’s lasting contribution to science was a single, uncomfortable idea: the system-level outcome is not reducible to the individual agents’ intentions. You can understand every agent perfectly — know its rules, its preferences, its decision process — and still be unable to predict what the system will do.

Symbolics, 1986: The Animator Who Made Birds Think

Craig Reynolds had no interest in economics or social science. He was a software engineer at Symbolics, a company that built graphics workstations, and he had a practical problem: he needed to animate realistic bird flocks for a short film.

The traditional animation approach — scripting each bird’s path — was hopeless. Real flocks have no choreographer. Hundreds of birds move as a coherent mass, splitting around obstacles and reforming, without any individual bird knowing the shape of the whole flock.

Reynolds tried something different. He gave each simulated bird (he called them “boids”) just three behavioral rules:

1. Separation — steer away from nearby flockmates to avoid collision
2. Alignment — steer toward the average heading of nearby flockmates
3. Cohesion — steer toward the average position of nearby flockmates

Each boid could only see its immediate neighbors. There was no central controller, no leader boid, no global awareness of the flock’s shape.

He presented the result at SIGGRAPH 1987. The boids flocked. They navigated obstacles, split into sub-groups, and reformed — all from three rules and local perception. The technique appeared in the 1987 short film Stanley and Stella in: Breaking the Ice, and a variant produced the bat swarms in Tim Burton’s Batman Returns (1992). In 1998, Reynolds received an Academy Scientific and Technical Award for his contributions to motion picture animation — three rules and an Oscar.

What Reynolds proved was a stronger version of Conway’s and Schelling’s insight: not only do simple rules produce complex behavior, but simple local rules can produce globally coherent behavior. The flock moves as one, not because anyone is coordinating it, but because each boid follows the same three rules based only on what it can see nearby.

The flip side was equally important: bad rules produce bad flocks. When Reynolds tweaked the separation distance, boids collided. When he weakened alignment, they scattered. The quality of the collective behavior was entirely a function of rule design, not agent intelligence.

Growing Artificial Societies

Joshua Epstein was a political scientist at Brookings who thought economics had an explanation problem. Economists could describe wealth inequality in a society — measure the Gini coefficient, plot the distribution — but they couldn’t generate it. They couldn’t start with individuals and simple rules and watch inequality appear. If you can’t grow it from the bottom up, Epstein argued, you don’t actually understand what causes it.

So he and mathematician Robert Axtell built a world to grow things in. They published Growing Artificial Societies in 1996. Their creation, Sugarscape, is a 51-by-51 grid where each cell contains some amount of sugar. Agents are placed on the grid. Each agent has vision (how far it can see), a metabolic rate (how much sugar it burns per turn), and a finite lifespan. The rules are minimal: look around, move to the richest visible cell, eat the sugar there.

They set up two peaks of sugar at opposite corners of the grid, hit run, and watched agents cluster around the peaks. Within a few hundred ticks, a skewed wealth distribution had appeared — a few agents with good vision and low metabolism had accumulated vast surpluses while others starved in sugar deserts. Nobody programmed inequality. It grew. When they added a second resource (spice) and allowed trading, markets formed with fluctuating prices. Cultural tags produced tribal groupings. Disease transmission created epidemics that swept through dense populations and skipped sparse ones.

The researchers could produce radically different artificial societies by changing nothing about the agents — same vision, same metabolism, same rules — and only changing the sugar distribution on the grid. Two peaks of sugar at opposite corners produced one kind of society. Sugar spread evenly produced a completely different one.

Epstein’s conclusion was direct: “If you didn’t grow it, you didn’t explain it.” Meaning: if you can’t generate a social phenomenon from the bottom up — from agent rules and environmental structure — then you don’t actually understand what causes it. Top-down explanations that start with the macro pattern and work backward are, in his view, just-so stories.

The Institute in the Desert

In 1983, George Cowan — a physicist who had worked on the Manhattan Project — started hosting lunches at Los Alamos for scientists who shared an unlikely suspicion: that the principles behind bird flocks, stock markets, immune systems, and urban sprawl might be the same principles. Over lunch, the conversation kept circling back to the same idea — that they needed an institute. Not government-run, not classified, genuinely open. Cowan recruited Nobel laureate Murray Gell-Mann, who later admitted that “many of the people I was calling were in very different fields and had never heard of me.”

The Santa Fe Institute opened in 1984. Its bet was that Conway’s cells, Schelling’s coins, Reynolds’ birds, and Epstein’s foragers were all instances of the same thing — complex adaptive systems, where autonomous agents interact in a shared environment and produce emergent behavior that no individual agent controls.

Under Chris Langton, SFI built Swarm, one of the first general-purpose platforms for agent-based simulation. Researchers used it to model financial markets, ant colonies, immune systems, and epidemics. By 2002, the National Academy of Sciences had convened a formal colloquium on the field.

Across thousands of studies and dozens of domains, two findings kept reappearing:

The environment shapes behavior more than agent intelligence does. Change the grid, the resource distribution, the network topology — and the same agents produce completely different outcomes. Smarter ants don’t make better colonies. Better pheromone trails do.

You cannot optimize the system by optimizing individual agents. The system’s behavior is an emergent property of agent-environment interaction. Improving one agent can worsen the whole. The only reliable lever is environment design.

January 2026: A Day in Gas Town

On January 15, 2026, Tim Sehn — co-founder of DoltHub, a company that builds a version-controlled database — tried Gas Town, Steve Yegge’s multi-agent orchestrator for Claude Code. Sehn pointed it at four failing tests in Dolt’s codebase and let the agents work.

Gas Town spun up twenty agents across twenty terminals, coordinated by a “Mayor” agent. Sehn described “a palpable sense of stress” watching the output stream by — too much happening at once to track. At one point the Mayor reported all four bugs were fixed. Only two pull requests existed on GitHub. Sehn corrected it; more agents spun up to investigate the gap.

Then one of the agents decided its work was done. It merged its own PR into the main branch. The integration tests were failing. Sehn couldn’t have caught it in time — by the time he saw the merge notification, broken code was already on main.

He shut Gas Town down and asked a colleague to force-push a reset. The sixty-minute session had burned roughly $100 in Claude tokens. “None of the PRs were good,” he wrote afterward, “and I ended up closing them all.”

I read Sehn’s post three times when it came out. What struck me wasn’t that the agents failed — it was how they failed. Not by writing bad code. By interacting with an environment that had no gate between “agent thinks it’s done” and “code reaches production.” No cap on autonomous merging. The agents weren’t broken. The landscape was.

Stripe’s “Minions” handle the same problem differently. Each Minion runs in an isolated devbox with a curated subset of 15 tools out of 400+ available. If tests fail twice, the task goes back to a human. No autonomous merging. They ship 1,300 PRs per week this way — but every one passes through a human review gate.

Same agents. Different environment. Different emergent behavior.

Conway’s cells, Schelling’s coins, Reynolds’ birds, Epstein’s foragers, Sehn’s coding agents, Stripe’s Minions — same mathematical structure. Autonomous agents following local rules in a shared environment, where the system-level outcome depends more on the environment than on the agents. They converge on the same solutions because the problem forces it — not because anyone read the papers.

Further Reading

If this made you want to go deeper, these are the books worth your time:

• Complexity: A Guided Tour — Melanie Mitchell. The best single introduction to the field. Covers cellular automata, genetic algorithms, information theory, and network science without requiring a math background. Start here.
• Growing Artificial Societies — Joshua Epstein & Robert Axtell. The Sugarscape book itself. Short, dense, and surprisingly readable. The appendix alone is worth it — they walk through every rule in the model.
• The Model Thinker — Scott E. Page. Practical framing for engineers: how to pick the right model for the problem you’re actually facing. Covers Schelling, Boids, network models, and twenty others in a single volume.
• Sync — Steven Strogatz. About spontaneous synchronization — fireflies, pacemaker cells, power grids, the Millennium Bridge. Same underlying math as emergence, told through a sequence of beautiful puzzles.
• Scale — Geoffrey West. West was president of the Santa Fe Institute. His central finding: biological organisms, cities, and companies all obey the same scaling laws — and the exponents reveal something deep about how complexity self-organizes.
• At Home in the Universe — Stuart Kauffman. The most speculative book on the list. Kauffman argues that self-organization is as fundamental as natural selection — that life and complexity don’t need a designer because the math makes them inevitable.