Agent-Based Modeling in the AI Age

Emergence from Simple Rules

Agent-based modeling (ABM) begins with a deceptively simple premise: instead of writing equations that describe a system from above, create agents with rules and let the system dynamics emerge from their interactions. No central planner designs the outcome. No equation predicts it. The macro-level pattern — segregation, wealth inequality, market crashes, epidemics — arises from the micro-level behavior of agents who cannot see the whole system and act only on local information.

This is the methodological counterpart to the complexity science foundations in Module 2. Where Module 2 explored how network structure shapes dynamics, ABM asks a different question: given agents with heterogeneous characteristics, adaptive behavior, and local interactions, what system-level phenomena emerge?

The canonical demonstration is Thomas Schelling Schelling segregation model Thomas Schelling’s 1971 model: households that only mildly prefer neighbours like themselves still end up in strongly segregated neighbourhoods. It is the classic demonstration of emergence. ’s segregation model (1971). Place two types of agents on a grid. Each agent has a mild preference: it wants at least 30% of its neighbors — 2 or 3 of its up-to-eight adjacent cells — to be its own type. An agent that falls short moves to a random empty cell; everyone else stays put. Notice what the rule does not ask for: no agent demands a majority of its own kind. Every agent is content living as a minority. Before you press Run, commit to a guess: once the grid settles, will it look well-mixed, mildly clumped, or starkly divided? And will settling take hundreds of steps, or fewer than ten?

You should have seen the segregation index climb from about 0.5 to roughly 0.73 within half a dozen steps — a well-mixed grid sorting itself into nearly homogeneous neighborhoods almost immediately. No agent wanted this outcome. No agent even preferred a majority of same-type neighbors. Yet the aggregate result is stark segregation, produced by a positive feedback loop Feedback loop A chain of cause and effect that loops back on itself. A reinforcing loop amplifies change (panic selling triggers more selling); a balancing loop damps it (a thermostat). : each move makes nearby agents of the other type slightly less satisfied, triggering further moves.

The natural next guess is that more intolerance means more segregation. Test it before reading on. The Strong preset (75%) does drive the index to near-total separation — about 0.99 — but watch the step counter: it takes well over a hundred steps, twenty times longer than the mild case. Now drag the threshold to 90% (the grid resets) and run again, committing to a prediction first. The result: segregation collapses. At 90%, almost no agent can ever be satisfied, so nearly everyone moves every step, and the churning grid stays mixed near 0.5 until the simulation gives up. More preference does not simply mean more separation — the relationship bends back on itself at the extreme.

This is the ABM insight in its purest form: the whole is not just more than the sum of its parts — it is qualitatively different. The behavior at the macro level (segregation) cannot be deduced from the behavior at the micro level (a mild preference) without running the simulation. The planets of Module 2 had a closed-form solution that made simulation a mere convenience; for this far simpler movement rule, no closed-form solution has ever been found. Only simulation reveals it.

One sign of how far the field has matured since Schelling computed his grids by hand: there are now documentation standards — the ODD protocol — that make models like this reproducible and comparable; the depth panel in the tool-landscape section below covers them.

Growing Artificial Societies

If Schelling demonstrates emergence Emergence Patterns of the whole that no single part intends or contains: a traffic jam from individual drivers, a price from individual trades. The whole behaves in ways you cannot read off the parts. from preference, Sugarscape demonstrates emergence from competition. Created by Joshua Epstein and Robert Axtell in their landmark 1996 book Growing Artificial Societies, Sugarscape places agents on a landscape of renewable resources (“sugar”) and gives them simple rules: look around, move to the cell with the most sugar within your vision range, harvest it, consume what your metabolism requires, and die if you run out.

The agents are heterogeneous — they differ in metabolism (1 to 4 units of sugar consumed per step) and vision (1 to 6 cells of sight). The landscape is uneven — two “sugar mountains” provide concentrated resources — and every agent starts with a random purse of sugar on a random cell. The simulator below tracks a single inequality number as the agents forage; for now, read it simply as higher means more unequal. Two experiments, and a committed guess before each. First, run the default heterogeneous world: does the number rise, fall, or wander? Second, click Equal Start — every agent now has identical metabolism and vision, keeping only its random birthplace and random starting purse. Does inequality still emerge?

In the heterogeneous world, the inequality number climbs from about 0.22 to roughly 0.38 by step fifty — and then locks in, plateauing rather than rising forever. Agents with wide vision and low metabolism accumulate reserves near the sugar peaks; others deplete theirs and die, the population falling from 150 to about 100. In the Equal Start world, the number falls — settling around 0.16, below where it began. Geography and starting luck wash out; it is the differences between agents that compound, step by step, into durable inequality. The number now deserves its name: it is the Gini coefficient, the standard measure of inequality, running from 0 (everyone owns the same) to 1 (one agent owns everything).

One caution about a word. “Distribution” here comes from statistics: it describes how frequently values occur — how wealth distributes itself across the population. It does not mean there is a distributor somewhere. No one carries out this distribution; no agent does anything to any other agent in this model. Each simply goes about its own foraging. To call the outcome an “unjust distribution” would miss the point — no agent is unjust. Yet the inequality is real, durable, and nobody’s decision.

That is Sugarscape’s importance for ABM methodology: it demonstrates how structural inequality can emerge from fair rules — a result that connects directly to Module 8’s complexity economics and the limitations of models that assume representative agents Representative agent A modelling shortcut that treats a whole population as one average decision-maker. It keeps the maths tractable but removes the differences between people that drive many crises. .

The ABM Tool Landscape

The tools have undergone a generational shift since 2015. NetLogo, the dominant platform for two decades, remains the entry point for education and rapid prototyping. Research and production have moved on: Mesa brought ABM into Python’s data-science and AI ecosystem, Agents.jl (Julia) serves models needing several times to two orders of magnitude more speed, and FLAME GPU 2 runs millions of agents on a single GPU Graphics processing unit (GPU) A chip built to do many simple calculations in parallel. Originally for video games, GPUs turned out to be ideal for training neural networks, and now dominate AI computing. workstation — simulations that previously required a computing cluster. Around them sit spatial platforms (GAMA), commercial tools (AnyLogic), and a shared model repository (COMSES.net).

That one-paragraph map is enough to follow the rest of this module — you may skip the panel below without losing the thread. If you want more, its Overview organizes the frameworks into four groups by use case, and the Detailed view adds architectures, benchmarks, the ODD documentation protocol, and the field’s venues and key figures.

AI-Powered ABM

Every simulation so far had a setting someone chose by hand: Schelling’s 30%, Sugarscape’s vision range. Real modelers usually face the reverse problem. They have data — the measured segregation of a real city, say — and need the settings that reproduce it. That search is called calibration, and done by brute force it means running the model thousands of times. AI is changing ABM in three ways: cheaper calibration, agents that learn their own rules, and models that can be tuned the way neural networks are trained. The third gets a demo you can run.

Surrogate models Surrogate model A fast stand-in for a slow simulation: a model (often a neural network) trained on a few hundred real runs to predict what the simulation would output, so thousands of settings can be tried on the stand-in instead. attack the cost of calibration directly. A surrogate is a fast stand-in for the real model: run the simulation a few hundred times, train a neural network (the pattern-learner of Module 10) to predict what it would output for any setting, then explore thousands of settings on the stand-in instead of the original. Once trained, the stand-in answers in a fraction of the time a full run takes, so the expensive model only has to be run a few hundred times. The search itself can be made smarter, too: Bayesian optimization Bayesian optimization A way to find good settings for a model that is expensive to run, using few tries: it keeps a rough statistical guess of how settings affect results and makes each next try where improvement looks most likely. chooses each next trial where the stand-in suggests improvement is most likely, instead of sampling settings blindly — which matters most when the model must match several real-world patterns at once.

Reinforcement learning Reinforcement learning (RL) A training regime where an agent learns by trial and error, nudged toward actions that lead to rewarded outcomes — which is why it proved itself first in games, where winning and losing supply the reward. replaces hand-written agent rules with learned ones. Instead of telling agents how to behave, the modeler rewards outcomes (wealth gained, distance saved, cooperation kept) and lets agents find strategies by trial and error. The Abmarl framework (Lawrence Livermore National Laboratory) connects ABM simulations to this kind of training for many agents at once, known as multi-agent reinforcement learning (MARL). This connects to Game Theory and Cooperation: MARL agents can learn to cooperate without being told how to communicate, and as conditions change, their group behavior can flip abruptly from coordinated to disordered — a phase transition Phase transition A sudden qualitative change produced by a smooth quantitative one: water becoming steam at 100°C, an epidemic taking off once transmission crosses a threshold, or an AI capability switching on at a certain scale. , like water turning to steam.

The third development needs one idea, and a demo supplies it. Most simulations are opaque to their own tuning: change a setting, rerun, and see what happens. A differentiable ABM Differentiable agent-based model An agent-based model written so a computer can calculate how each output changes when each setting is nudged. That lets it be fitted to data with the same gradient methods used to train neural networks. is written so the computer can also report which direction each setting should move to bring the output closer to the data, and how steeply the output responds. That turns calibration from a blind search into a walk downhill. Frameworks such as AgentTorch and its full write-up, FLAME (MIT Media Lab with Adobe; a NeurIPS 2023 workshop paper and AAMAS 2024), are built to supply that direction automatically, for millions of agents on a single GPU; the depth panel after the demo says how.

The demo below turns that idea from a claim into a count. The task: find the Schelling threshold that produces a chosen target segregation level. Both strategies may only run the model and observe the result — no formula, no shortcut. Random search will spend exactly 40 runs. The other strategy follows the slope downhill (its button is labelled Gradient Descent; the name is explained just after). Commit to a guess before clicking: how many runs will it need to land near the target — more than 40, about the same, or far fewer?

Count the evaluations: at the default target, gradient descent typically lands within tolerance after 7 to 16 model runs, against random search’s fixed 40 — the counter under each button is the checkable number. Now the vocabulary for what you just watched. The grey curve — how far each threshold lands from the target — is a loss function Loss function A formula that scores how wrong a model’s outputs are, so that fitting the model means making that score as small as possible. . Its slope at any point is the gradient, and walking downhill along that slope is gradient descent Gradient descent The standard way of fitting a model: find the direction of adjustment that reduces the error fastest, take a small step that way, and repeat. It is like walking downhill in fog by feeling the slope. . Random search ignores the slope and pays for it in evaluations; gradient descent reads the slope and spends its budget only where the curve points. Differentiable ABM frameworks exist to make that slope available automatically, even for models with millions of agents and parameters.

Scaled up, that difference is the case for the method. GradABM, an early differentiable epidemic model, simulates million-agent populations in a few seconds on ordinary hardware, which puts gradient-based calibration within reach of models too large for brute-force search; and sensitivity analysis — how much each setting matters — comes out of the same calculation instead of thousands of extra runs.

That is the practical picture, and it is enough for the rest of the module — the panel below is optional depth. Its Overview explains how frameworks keep randomness differentiable (the reparameterization trick); the Detailed view covers AgentTorch’s tensor representation and hybrid learning modes.

LLM Agents and the Validation Challenge

The most dramatic recent development is using large language models as agent cognition. Instead of hand-coded rules or learned RL policies, LLM Large language model (LLM) An AI model trained on vast amounts of text to predict the next word-piece. Chatbots such as ChatGPT and Claude are built on them. -powered agents perceive their environment through text descriptions, reason using the LLM’s capabilities, and produce actions in natural language. Stanford’s “Generative Agents” paper (2023) demonstrated LLM agents planning daily routines, forming relationships, and organizing social events in a simulated town — behavior far more human-like than any rule-based agent could produce.

LLM-assisted ABM creation is equally transformative. A CHI 2024 study of NetLogo Chat, an LLM assistant for writing agent-based models, found that experienced modelers got more out of it than novices; other work describes multi-stage workflows in which the LLM translates natural language specifications into Mesa or NetLogo code, accelerating the model development cycle from weeks to hours. The Tsinghua FIB-Lab’s survey (Humanities and Social Sciences Communications, a Nature Portfolio journal, 2024) mapped the full landscape of LLM-ABM integration.

But this power comes with a fundamental challenge: validation. As Larooij and Törnberg (2025) argue in Artificial Intelligence Review, LLM integration may exacerbate rather than alleviate ABM’s validation crisis — the field’s long-standing difficulty in showing that a model matches the world rather than merely producing plausible output. The core problems are interconnected:

Black-box behavior: LLM agents make decisions through mechanisms that are opaque to the modeler. Traditional ABM rules are transparent — you can trace exactly why an agent moved. LLM agent decisions pass through billions of parameters with no interpretable causal chain.

Reproducibility: LLMs are stochastic — the same prompt can produce different outputs across runs, API versions, and even temperature settings (the dial that sets how much randomness the model adds to its answers). This undermines the reproducibility that the ODD protocol was designed to ensure.

Bias inheritance: LLM agents inherit the biases of their training data. A simulated population of LLM agents may exhibit WEIRD (Western, Educated, Industrialized, Rich, Democratic) biases rather than the behaviors of the population being modeled.

Scalability: Running an LLM query per agent per step is computationally expensive. Simulations with thousands of LLM agents face significant GPU and API cost constraints that traditional ABMs do not.

This creates a methodological fork: mechanistic ABMs offer transparency and reproducibility but limited behavioral realism; LLM-powered ABMs offer rich, human-like behavior but at the cost of interpretability and validation rigor. The field has not yet resolved this tension.

How would anyone settle it? Modelers grade validation in three tiers, each stricter than the one before. This site uses these three throughout. Take the Schelling grid from the top of this module:

• Face validity asks whether the output looks right to someone who knows the domain. Do the clusters resemble real neighborhoods?
• Behavioral validity asks whether the model’s parts and patterns match measurements. Do real households move at thresholds like the agents’ — and does the simulated segregation index match the one measured for real cities?
• Predictive validity asks whether the model forecasts what it was not fitted to. Given a city’s mix today, does it say how its neighborhoods will look in ten years, before the decade has passed?

Face validity is cheap, and LLM agents pass it almost by construction: they are trained to produce plausible behavior. That is why the fork bites. The tier LLM agents reach most easily is the one that proves least. Module 15 checks how far today’s LLM-agent studies have climbed.

Meanwhile, evolutionary game theory — building on Game Theory and Cooperation — has spread fast into applied modeling since 2019: a bibliometric review of its use in healthcare alone counted 539 publications from 2000 to mid-2025 (tumor heterogeneity as evolutionary games, vaccine logistics as tripartite games), and similar applications reach sustainability (fisheries and water management as N-player commons dilemmas, smart grid economics as evolutionary games against dynamic pricing) and governance (stakeholder strategy co-evolution). These applications demonstrate ABM’s value precisely where validation is most critical — in domains where policy decisions affect millions.

If the fork’s practical consequences are all you need, skip the panel below and move on to the applications — nothing later depends on it. For readers who want the machinery: the Overview recaps how validation was done before LLM agents, and the Detailed view adds the formal verification–validation–accreditation framework and the economics of LLM-agent scale.

Applications and the Road Ahead

ABM has moved from academic exploration to operational deployment across multiple domains. COVID-19 was a watershed: individual-based models like Covasim captured heterogeneous contact patterns, superspreading events, and the effects of targeted interventions (school closures, workplace policies, vaccination prioritization) that aggregate compartmental models like SIR could not resolve. The pandemic demonstrated ABM’s value for policy under uncertainty — and its limitations when calibration data was sparse and fast-changing, as explored in Module 7’s complexity lens on COVID.

Smart cities and urban planning use ABM to simulate traffic, pedestrian flow, energy systems, and land use change. GAMA and MATSim power large-scale urban simulations integrating GIS data with agent behavior. Climate applications model farmers’ adaptation to changing conditions, migration patterns driven by environmental stress, and the cascading effects of extreme weather through interconnected infrastructure — connecting to Module 2’s resilience analysis.

Financial markets — the domain explored in Module 8’s market simulator — use ABM to model heterogeneous trader strategies, flash crashes, and systemic risk. Central banks including the Bank of England, ECB, and Federal Reserve now use ABM alongside traditional DSGE Dynamic stochastic general equilibrium model (DSGE) The standard macroeconomic model used by central banks since the 2000s. It treats the economy as a few representative households and firms that optimize and settle into equilibrium, with random shocks pushing them around. models for policy simulation.

The computational frontier continues to advance. A 2021 study ran a full-scale model of an economy the size of the eurozone — 331 million agents — in 108 seconds on 128 CPU cores, at about 70% strong-scaling efficiency. Generative social simulations have exceeded 10,000 LLM-powered agents with some 5 million interactions per run. Digital twins Digital twin A continuously updated computer model of one specific real thing, such as a jet engine, a factory or a city, used to monitor it and to test changes before making them. — live ABMs continuously calibrated against real-time data — represent the convergence of ABM, IoT, and AI infrastructure; Module 15 develops them in full.

The field’s trajectory points toward a synthesis: differentiable ABM cores providing gradient-based optimization, GPU acceleration enabling million-agent scale, surrogate models making calibration tractable, and LLM agents providing behavioral richness where needed — all within reproducible, ODD-documented frameworks. The gap between this vision and current practice remains significant. But the tools, the community, and the demonstrated applications make agent-based modeling the most natural computational methodology for the complex systems that define our world.