In the weeks before a rebellion, a city can look calm. People go to work, queue at markets, nod to police at checkpoints. Ask any one of them what they think of the regime and you might hear something far angrier than anything visible on the street. Then, one day, seemingly out of nowhere, a crowd forms, and within hours a neighbourhood that looked pacified for a decade is in open revolt.
Analysts explaining the moment afterward tend to reach for a trigger like a single killing or a fuel price hike. There usually is one, but the trigger is rarely the interesting part. The interesting part is why a population that seemed compliant a week earlier turned out to be sitting on a reservoir of grievance nobody had measured, and why the same trigger, a year earlier or a year later, might have produced nothing at all.
Most strategic analysis explains moments like this with a linear story: grievance built up, a trigger arrived, the population responded. It’s an intuitive shape for a narrative, and it fits neatly into a policy brief. But it treats a population as a single actor with a single grievance level, and it treats the outcome (revolt or no revolt) as a fairly direct function of how bad things had gotten. If the linear story were right, you could read a population’s mood, plot its trajectory, and get a reasonable forecast of when it boils over.
That’s roughly the assumption behind the tools political science has historically reached for to study conflict: statistical regressions relating measured national indicators to conflict onset, or formal game-theoretic models that solve for a single equilibrium outcome given a specified set of payoffs. Both are useful. Both also require something the linear story needs to be true which is that the relevant unit of analysis behaves roughly as one actor, and that history unfolds close to its expected path (Masad, 2016, pp. 1–3). Agent-based modelling exists because that assumption usually isn’t true, and because the actual texture of political interaction which is actors watching each other, copying each other, retaliating, changing their minds when the person next to them changes theirs, disappears the moment you average everyone into a single number.
Agent-based modelling starts from the opposite assumption. Instead of one actor with one grievance level, you build a population of individuals, each with their own circumstances, their own tolerance for risk, and crucially only a local view of what everyone around them is doing. Each one follows a simple rule for how to behave given what it can see. Then you run the simulation and watch what happens at the level of the whole system.
This sounds like a modest change but it isn’t. Once agents are heterogeneous, locally informed, and reactive to each other rather than to a global average, the link between individual circumstance and collective outcome stops being linear and in some cases stops being predictable from the input parameters at all. Lars-Erik Cederman’s GeoSim model, built to explain why wars follow a power-law size distribution found that this pattern emerges from roughly two hundred interacting, spatially situated states without any actor or mechanism in the model aiming to produce a power law. It’s a signature of what physicists call self-organised criticality: a system that organises itself, through purely local interactions, into a state where a small disturbance can occasionally cascade into a very large one, for reasons that have nothing to do with the size of the disturbance itself (Cederman, 2003, pp. 135–137). You don’t get that kind of finding from a model that treats the international system as one actor solving one equation. You need agents reacting to their neighbours.
The same logic runs through a wider range of agent-based work in international relations. Bearce and Fisher’s (2002) model of trade and war makes the point explicitly: conventional econometric approaches “treat either trade or war as exogenous, ignoring that both are emergent, endogenous processes influenced by the international system and state behaviour” (p. 365). Their model lets trading networks and war both emerge from the same underlying state-level interactions, rather than assuming one causes the other from outside the system. Robert Axelrod’s work on the evolution of cooperation makes a related point from the opposite direction: strategies like tit-for-tat, and the social norms that stabilise cooperation once punishment becomes decentralised and self-enforcing, were not designed by anyone emerged from thousands of local, myopic interactions among simple agents (Axelrod, 1998, pp. 40–68). Ian Lustick’s ABIR model of identity formation found something similarly non-obvious: give agents a moderate repertoire of possible identities rather than a single fixed one, and inter-group tension drops sharply but push the repertoire size too far and the population fragments instead of consolidating, a curvilinear relationship nobody wrote into the model’s rules on purpose (Lustick, 2000).
What unites these examples is a shared method: specify how individual actors behave locally, let the interactions run, and treat whatever pattern shows up at the system level as something to be explained rather than assumed. That shift from modelling outcomes directly to modelling the interactions that produce them is what agent-based modelling actually offers strategic analysis. It isn’t a better crystal ball but it’s a different question.
Nowhere is this clearer than in Joshua Epstein’s (2002) model of civil violence, still one of the most cited agent-based models in political science two decades on, precisely because it is so simple and so revealing. Epstein’s isn’t a model of any specific uprising. It’s a stylised population scattered across a grid, each agent carrying two properties: a level of hardship, and a level of risk aversion. Each agent’s grievance is a simple function of its hardship and its view of the regime’s legitimacy. High legitimacy suppresses grievance even under real hardship; low legitimacy inflames it (Epstein, 2002, p. 7243). Police agents patrol the same grid and arrest active rebels they can see. Every agent’s decision to openly rebel or stay quiet comes down to one comparison: is my grievance bigger than my locally perceived risk of arrest, estimated by looking at how many police versus how many other rebels I can currently see around me?
That’s the entire rule set. No agent has a theory of revolution. No agent is trying to organise anything. And yet, run over enough cycles, this population reliably produces patterns that look uncannily like real political upheaval which are patterns Epstein never told the model to produce.
The first is deception. Highly aggrieved agents surrounded by police stay quiet, not because their grievance has changed but because their locally perceived arrest risk is high. The moment police move on, the same agents turn openly rebellious. Nothing in the model represents “concealment” as a strategy; it falls straight out of the arithmetic of the activation rule. A society under this model can look completely pacified while harbouring grievance that surfaces the instant the visible threat lifts (Epstein, 2002, p. 7245).
The second is cascade. If a few agents happen to cluster somewhere with few police nearby, the local police-to-rebel ratio drops, which lowers the perceived risk for everyone nearby, which draws in agents previously too cautious to act, which lowers the ratio further still. This is why authoritarian regimes so consistently restrict assembly: the model shows, mechanistically, why a crowd is dangerous to a regime in a way the same number of individually aggrieved but dispersed people is not (Epstein, 2002, p. 7245).
The third is the least intuitive and the most useful for an analyst. Epstein compared what happens when a regime’s legitimacy erodes gradually against what happens when it drops suddenly by the same total amount. Gradual erosion barely moves the needle. The most aggrieved individuals cross their threshold one at a time and get arrested before enough of them cluster to trigger a cascade. A sudden shock of the same total size does the opposite: many agents cross the threshold simultaneously, overwhelming local policing before it can respond, and the cascade dynamic takes over. Epstein calls the gradual version a “salami tactic” — a regime can lose an enormous amount of legitimacy, sliced thin enough, and never face a large uprising for it (Epstein, 2002, p. 7247). The rate of change, not the total change, is what matters. A parallel result holds for repression: gradual liberalisation, not gradual crackdown, is what tends to push the system past its tipping point which is a computational echo of Tocqueville’s observation that revolutions tend to arrive not at the peak of repression but just as it starts to ease (Epstein, 2002, pp. 7247–7248).
None of this was programmed in directly as those are outcomes of the model, not inputs to it, which is exactly the distinction the whole method depends on. Epstein’s own extension of the model to interethnic violence adds one more result worth sitting with: when he varied the density of peacekeeping forces meant to prevent escalation into genocide, more peacekeepers did, on average, buy more time before catastrophe but the variance around that average grew just as fast as the average itself. At every force level he tested, some runs collapsed into genocide almost immediately and others held for the full length of the simulation. Force size alone did not predict which (Epstein, 2002, pp. 7249–7250).
None of this tells an analyst when the next uprising will happen. That’s precisely the point, and precisely the value. What Epstein’s model offers instead is a different set of questions to ask about a population that looks stable.
First: stability is not evidence of consent. A population showing no visible unrest can still be sitting on concealed grievance that a change in visible enforcement (a troop withdrawal, a policing gap, a moment of regime distraction) could expose all at once. Second: watch the rate of change in the signals that matter, not just their level. A slow, sustained erosion of legitimacy or living standards may be more survivable for a regime than a single sharp shock a fraction of the size, because gradual decline lets a state pick off would-be catalysts one at a time before they cluster. Third: be suspicious of any claim that a fixed intervention size (a peacekeeping force, a troop surge, a sanctions package) guarantees a predictable outcome. Epstein’s model shows the same intervention, at the same scale, producing wildly different timelines to catastrophe depending on essentially where and when local clustering happens to occur. An analyst briefing a decision-maker on force sizing should be honest that more generally helps on average, while being explicit that “on average” is doing a lot of work in that sentence.
None of this should be mistaken for a validated theory of revolution. Epstein is explicit that hardship, legitimacy, and risk aversion in his model are stylised constructs, not measured quantities. The model is a thought experiment about mechanism, not a calibrated forecasting tool (Epstein, 2002, p. 7250). That caution generalises across the method. Agent-based models are only as good as the behavioural rules built into them, and those rules are frequently unvalidated assumptions dressed up in the language of a formal model. Even ambitious, well-resourced attempts to bring empirical rigour to agent-based conflict models run into this wall: Devon Masad’s (2016) attempt to reimplement and validate Bruce Bueno de Mesquita’s widely cited Expected Utility Model (a model that has actually been used for real policy forecasting) could not fully reproduce the original’s own published results, in part because the original was never fully specified or made transparent (Masad, 2016, pp. 94–101). If a model already in policy use can’t be independently verified against its own claimed outputs, that’s a serious caution against treating any single agent-based model’s results as settled fact.
Cederman and Girardin (2023), reflecting on their own decades of work in this field, make the same point from the inside: the abstraction that gives agent-based models their explanatory power is also what limits their empirical validation and their influence on policy which is part of why the field shifted toward pairing ABM with spatial, empirically grounded data in the 2000s rather than relying on pure simulation alone. The question to ask of a model like Epstein’s isn’t whether to trust it outright, but the same question you’d ask of a single dramatic case study: is this one plausible mechanism among several, or the mechanism? Usually it’s the former which is exactly what makes it worth thinking with, and exactly why it should never be the only thing you think with.
Go back to the city that looked calm the week before it wasn’t. The linear story says the trigger caused the revolt. The interaction story says something closer to the truth: a population of people, each reading the risk on their own street corner, had been sitting one shock away from a cascade for longer than anyone watching from outside could have known. The trigger didn’t create the grievance. It just told everyone, all at once, that the street corner next to them was ready too.
The next piece in this series looks at what’s actually inside an agent like the ones in Epstein’s model: the decision rule itself, and how bounded rationality rather than not full optimisation, is what makes an agent like this behave the way it does. That’s next: Bounded Rationality and Political Agents.
Further reading
Axelrod, R. (1998). The Complexity of Cooperation: Agent-Based Models of Competition and Collaboration. Princeton University Press.
Bearce, D. H., & Fisher, E. O. N. (2002). Economic geography, trade, and war. Journal of Conflict Resolution, 46(3), 365–393.
Cederman, L.-E. (2003). Modeling the size of wars: From billiard balls to sandpiles. American Political Science Review, 97(1), 135–150.
Cederman, L.-E., & Girardin, L. (2023). Computational approaches to conflict research: From modelling and data to computational diplomacy. Journal of Computational Science.
Epstein, J. M. (2002). Modeling civil violence: An agent-based computational approach. Proceedings of the National Academy of Sciences, 99(Suppl 3), 7243–7250.
Lustick, I. S. (2000). Agent-based modelling of collective identity: Testing constructivist theory.
Masad, D. P. (2016). Agents in Conflict: Comparative Agent-Based Modelling of International Crises and Conflict [Doctoral dissertation, George Mason University].