Skip to main content
Category: Models
Type: Systems / computational simulation model
Origin: Thomas C. Schelling (1971); Joshua M. Epstein and Robert Axtell (1996)
Also known as: Multi-agent simulation; individual-based model; agent-based computational model
Quick Answer — An Agent-Based Model is a simulation in which many autonomous, often heterogeneous agents follow local rules, interact with one another and with an environment, and thereby generate system-level patterns. Thomas Schelling’s 1971 segregation model showed the logic with coins on a checkerboard; Joshua Epstein and Robert Axtell’s 1996 Sugarscape made it a laboratory for “growing” societies. The practical lesson is to ask whether a macro pattern can be grown from honest micro rules—before you infer motives from the aggregate, or average everyone into one representative actor.

​
What is an Agent-Based Model?

An Agent-Based Model is a computational experiment: you specify agents, the space they inhabit, and the rules they follow, then watch whether a recognizable large-scale pattern appears from the bottom up.
Perhaps one day people will interpret the question, “Can you explain it?” as asking “Can you grow it?”
Picture a cafeteria at noon. Nobody is assigned a table. Each person only wants a few familiar faces nearby. After a handful of seat changes, the room has split into clumps that no one planned. That is the everyday cue: the map of tables is not a faithful portrait of how strongly anyone wanted to exclude others. Systems thinking looks at the whole; the Agent-Based Model is the sandbox that lets the whole arise from parts that are allowed to differ. It is a complement to a stock and flow picture, which tracks totals, not the distinct people who make those totals move.

​
Agent-Based Model in 3 Depths

  • Beginner: When a crowd, market, or neighborhood surprises you, ask what each person could see and do locally—not only what the average person wanted.
  • Practitioner: Name the agents, their states, the environment, and one or two interaction rules. Run the model, then change one rule and see which macro pattern survives.
  • Advanced: Growing a pattern shows sufficiency, not unique cause. Calibration, stochastic runs, and an honest description of what the agents cannot see are what keep the sandbox from becoming a movie.

​
Origin

The Agent-Based Model has several ancestors, not a single inventor. In the 1950s, John von Neumann and Stanislaw Ulam explored cellular automata: cells on a grid that update from their neighbors. That lineage later included Conway’s Game of Life and the edge of chaos studies of the 1980s. In 1987, Craig Reynolds’s SIGGRAPH paper “Flocks, Herds, and Schools: A Distributed Behavioral Model” showed that three local steering rules—separation, alignment, cohesion—could produce flocking without a leader. The social-science landmark is earlier and more austere. Thomas C. Schelling, then at Harvard and writing for RAND, issued Models of Segregation as RM-6014 in May 1969, and published “Dynamic Models of Segregation” in the Journal of Mathematical Sociology (1971, Vol. 1, No. 2, pp. 143–186). He moved pennies and dimes on a checkerboard: two recognizable groups, some vacant cells, and a mild wish not to be a local minority. Extreme clustering followed. He later restated the argument in Micromotives and Macrobehavior (1978). The Nobel Prize in Economic Sciences (2005, shared with Robert Aumann) cited his work on conflict and cooperation; the segregation model remains the teaching case for micromotives versus macrobehavior. In 1996, Joshua M. Epstein and Robert Axtell, working at the Brookings Institution and connected to the Santa Fe Institute, published Growing Artificial Societies: Social Science from the Bottom Up (MIT Press). Their Sugarscape placed agents on a twin-peaked resource grid. A local foraging rule—look as far as your vision allows, move to the richest nearby cell, eat, metabolize—was enough to grow migration, trade, and a skewed wealth distribution. Epstein’s 1999 essay “Agent-Based Computational Models and Generative Social Science” in Complexity (Vol. 4, No. 5, pp. 41–60) named the method: explanation as generative sufficiency. Two later papers made the craft more teachable. Eric Bonabeau’s 2002 Proceedings of the National Academy of Sciences survey (Vol. 99, Suppl. 3, pp. 7280–7287) set out when the method beats a representative-agent equation. Volker Grimm and 27 coauthors proposed the ODD protocol—Overview, Design concepts, Details—in Ecological Modelling (2006, Vol. 198, pp. 115–126), so that individual-based and agent-based models could be read and replicated instead of remaining black boxes.

​
Key Points

Use an Agent-Based Model when the people (or firms, cells, vehicles) are not interchangeable, and the pattern you care about lives in their interaction.
1

Give agents states, not just a group average

Agents can differ in vision, patience, wealth, location, or memory. A single “representative” person cannot produce clustering, queues, or cascades that depend on who sits next to whom. If your spreadsheet already treats everyone as the mean, you are not yet doing this kind of model.
2

Write local rules before you chase a global target

Each agent sees a neighborhood, a price, or a few neighbors—not the whole system. Schelling’s movers asked only whether enough nearby seats felt familiar. Reynolds’s boids aligned with nearby birds. If a rule needs a god’s-eye dashboard, it is no longer a local rule, and you will smuggle the conclusion into the setup.
3

Treat emergence as a result to be grown, not a slogan

A macro pattern—segregation, a traffic jam, a tipping point, a fat tail of wealth—counts as grown when it appears from the stated rules. Epstein’s test is strict: if you cannot grow it, you have not shown that those microspecifications are even sufficient. Growing it still does not prove they are the true cause.
4

Document what can be replicated

Grimm’s ODD checklist asks for purpose, entities and scales, process scheduling, initialization, inputs, and submodels. Without that, a colorful run is an anecdote. Pair the story with repeated runs, as you would with Monte Carlo simulation, because many agent models are stochastic.

​
Applications

The model earns its keep when a local decision, repeated by unlike actors, could create a pattern that no one chose.

Neighborhoods, classrooms, and offices

Ask what each household, student, or team can see, then let them move or regroup. If mild “enough people like us” rules already split the map, do not jump to a story of extreme preference. The same sketch diagnoses self-sorting in lunch tables and in org charts.

Traffic, crowds, and queues

Give each driver or pedestrian a following distance and a reaction time. Jams can form from braking chains with no accident and no villain. Change one local rule—merge etiquette, door placement, batch size—before you rebuild the whole corridor.

Contagion of disease, rumor, or products

Average contact rates hide hubs, households, and workplaces. Heterogeneous agents let a few busy nodes dominate spread, which a well-mixed equation will miss. Use the model to ask which contact to cut, not to mint a false precise peak.

Markets, commons, and resource use

Sugarscape-style foraging, or a tragedy of the commons with unlike users, can grow inequality or collapse from simple metabolism and vision. Stress-test a tax, a quota, or a sharing rule by changing it in the sandbox before you argue from a single average farmer.

​
Case Study

Thomas Schelling’s checkerboard is still the cleanest demonstration that micromotives need not match macrobehavior. In the 1969 RAND memorandum and the 1971 journal article, two groups occupy a finite grid. About 25–30% of cells are vacant, so movement is possible. Each household looks at a local neighborhood—in the spatial version, often the 5 × 5 block around it—and relocates if too many neighbors are of the other type. The measurable surprise is the gap between the demand and the outcome. Schelling reported that once the process settles, the effective preference realized in the typical neighborhood sits near one-third similar neighbors, and that segregation rises steeply as demands move through the 35–50% range. Households that would have tolerated mixed blocks still end up in exaggerated clumps. His own summary is the finding: there is no simple correspondence of individual incentive to collective results, and you usually cannot read individual motives off the aggregate pattern. That is why the model belongs in an Agent-Based Model article rather than only in a prejudice article. The engine is local movement plus a threshold, not a central command to separate. Boundary note: real cities also run on prices, lending rules, violence, and formal discrimination. Growing segregation from mild preference shows a possible mechanism. It does not prove that mechanism is the whole history of any actual city.

​
Boundaries and Failure Modes

An Agent-Based Model shows what can happen, not what must have happened. Many different micro-rules can grow the same curve. If you stop at “we reproduced the shape,” you have a candidate, not a verdict. That is a sufficiency trap. It is also easy to overfit. Extra agent traits, extra networks, and extra stochastic knobs will fit last year’s data and fail next year’s. Bonabeau’s practical warning still holds: use agents when heterogeneity, local interaction, or adaptive rules are the point; otherwise a stock and flow or equation model is clearer and cheaper. Do not treat a prettier animation as more scientific. The common misuse is motive inference. A segregated map, a jammed freeway, or a polarized feed gets read as proof that “people wanted this.” Schelling’s point, and Epstein’s, is the opposite: check the local rules and the feedback loops they close before you indict the average heart.

​
Common Misconceptions

The method looks like a movie, so it is easy to confuse animation with explanation, and explanation with proof.
Cellular automata are one ancestor. A useful Agent-Based Model states who the agents are, what they can observe, how they update, and which macro pattern would count as success or failure. A pretty grid without those claims is a demo, not a model.
Growing a pattern shows that those rules are sufficient in silico. Real systems may use different rules that produce the same shape. Treat the run as a candidate mechanism, then look for independent traces of the micro-rules.
Extra traits explode the things you must justify and the runs you must check. A small set of heterogeneous rules that you can document in ODD form usually teaches more than a crowd of guessed psychologies.
These pages sit next to the same problem: how local action and system structure produce patterns that no single actor chose.

Systems Thinking

Seeing relations and delays before you blame a single actor.

Stock and Flow

The aggregate picture of levels and rates that agent models can generate but need not start from.

Feedback Loops

How local moves close reinforcing or balancing loops at system scale.

Tipping Point Model

Threshold language for when local dissatisfaction triggers a cascade of moves.

Monte Carlo Simulation

Repeated random runs that a stochastic Agent-Based Model still needs for a range, not one movie.

Computational Thinking

The habit of turning a process into states, rules, and an experiment you can rerun.

​
One-Line Takeaway

Write the local rules honestly, try to grow the pattern, and only then decide what the crowd “must have wanted.”