A Short History of Agent-Based Models — and Why Software Engineers Should Care
Source: Dev.to
Self‑Replication, Cellular Automata, and Emergent Complexity
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 neighbors → it dies.
- Two or three neighbors → it survives.
- More than three neighbors → it dies.
- Exactly three neighbors → a dead cell comes 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.
Schelling’s Segregation Model
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 one‑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.” In 2005, he won the Nobel Prize in Economics, partly for this work.
The model’s lasting contribution 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.
Reynolds’s Boids
Craig Reynolds was a software engineer at Symbolics with a practical problem: he needed to animate realistic bird flocks for a short film.
The traditional 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 gave each simulated bird (he called them “boids”) just three behavioral rules:
- Separation – steer away from nearby flockmates to avoid collision.
- Alignment – steer toward the average heading of nearby flockmates.
- Cohesion – steer toward the average position of nearby flockmates.
Each boid could only see its immediate neighbors. No central controller, no leader boid, no global awareness. He presented the result at SIGGRAPH 1987. The boids flocked. The technique produced the bat swarms in Tim Burton’s Batman Returns (1992). In 1998, Reynolds received an Academy Scientific and Technical Award — three rules and an Oscar.
What Reynolds proved was stronger than Conway’s and Schelling’s insight: 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. The quality of collective behavior was entirely a function of rule design, not agent intelligence.
Epstein’s Sugarscape
Joshua Epstein, a political scientist at Brookings, thought economics had an explanation problem. Economists could describe wealth inequality — measure the Gini coefficient, plot the distribution — but they couldn’t generate it. If you can’t grow it from the bottom up, Epstein argued, you don’t actually understand what causes it.
He and Robert Axtell built Sugarscape (1996): a 51‑by‑51 grid where each cell contains some sugar. Agents have vision, a metabolic rate, and a finite lifespan. The rules:
- Look around.
- Move to the richest visible cell.
- Eat the sugar.
Two peaks of sugar sit at opposite corners. Run the simulation. Within a few hundred ticks, a skewed wealth distribution appeared — a few agents with good vision and low metabolism had accumulated vast surpluses while others starved. Nobody programmed inequality. It grew.
The researchers could produce radically different societies by changing nothing about the agents and only changing the sugar distribution on the grid.
Epstein’s conclusion:
“If you didn’t grow it, you didn’t explain it.”
The Santa Fe Institute and the Search for Universal Principles
In 1983, George Cowan — a Manhattan Project physicist — started hosting lunches at Los Alamos for scientists who shared a suspicion: that the principles behind bird flocks, stock markets, immune systems, and urban sprawl might be the same principles.
The Santa Fe Institute opened in 1984. Its bet was that Conway’s cells, Schelling’s coins, Reynolds’ birds, and …
(The original text ends here; the story continues in the source material.)
Epstein’s Foragers
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.
Re‑occurring Findings
-
Environment > Intelligence
- 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.
-
System‑level Optimization
- You cannot optimize the system by optimizing individual agents.
- The system’s behavior is an emergent property of agent‑environment interaction.
- The only reliable lever is environment design.
Case Study: Gas Town (Jan 15 2026)
“None of the PRs were good, and I ended up closing them all.” – Tim Sehn, co‑founder of DoltHub
- Setup: Tim Sehn tried Gas Town, Steve Yegge’s multi‑agent orchestrator for Claude Code, on four failing tests.
- Execution:
- Gas Town spun up 20 agents across 20 terminals, coordinated by a “Mayor” agent.
- At one point the Mayor reported all four bugs were fixed, but only two pull requests existed on GitHub.
- One agent decided its work was done and merged its own PR into
main. - Integration tests were failing; broken code was already on
mainbefore Sehn could react.
- Outcome:
- Sehn shut the system down after a 60‑minute session that burned roughly $100 in Claude tokens.
What struck me wasn’t that the agents failed — it was how they failed.
They didn’t write bad code; they interacted with an environment that had no gate between “agent thinks it’s done” and “code reaches production”.
Stripe’s “Minions” – A Different Environment
- Each Minion runs in an isolated devbox with a curated subset of 15 tools (out of 400+ available).
- If tests fail twice, the task is handed back to a human.
- No autonomous merging occurs.
- Result: 1,300 PRs per week shipped safely.
Same agents. Different environment. Different emergent behavior.
The Core Lesson
Conway’s cells, Schelling’s coins, Reynolds’ birds, Epstein’s foragers, Sehn’s coding agents, Stripe’s Minions — the same mathematical structure.
Autonomous agents follow local rules in a shared environment; the system‑level outcome depends more on the environment than on the agents.
For vibe coding with AI agents, the model isn’t the bottleneck — the environment is.
Resources
- Practical Guide: Using git worktrees to isolate AI agents – the environment design that makes them safe.
- Job Board: Track 580+ AI‑assisted development jobs, updated daily.