Agent-Based Modelling

Agent-based modelling (ABM) is a computational paradigm in which systems are represented not through aggregate equations but through populations of autonomous, heterogeneous agents interacting according to local rules. The goal is not to simulate a system in its totality but to grow it from the bottom up: to plant a set of agents in an environment, give them rules, and observe what structures emerge that no individual rule commands. ABM is the methodological commitment that macroscopic regularity is not assumed but produced.

The Core Commitment

An agent-based model contains three essential elements: agents with internal states and decision rules, environments that constrain and enable action, and interaction topologies that determine who can affect whom. Unlike equation-based models, which begin by assuming the macroscopic pattern (a differential equation for population growth, a market-clearing price), ABM begins by assuming the microscopic rules and discovers the macroscopic pattern as an emergent consequence. This makes ABM particularly suited to phenomena where heterogeneity matters: financial markets with diverse trader strategies, epidemics where contact networks structure transmission, or ecosystems where individual foraging decisions reshape the landscape.

The ontological wager of ABM is that the right level of description for many social and biological phenomena is the individual, not the aggregate. This is not always true — fluid dynamics is better handled by partial differential equations than by simulating individual molecules — but ABM practitioners claim it is true more often than classical methods assume. The field of Agent-Based Economics has demonstrated that markets with heterogeneous agents can produce bubbles, crashes, and persistent inequality that representative-agent models cannot capture.

Domains and Applications

ABM has become a lingua franca across disciplines precisely because the interaction of heterogeneous agents is a universal pattern. In epidemic modelling, ABM captures the non-random mixing that determines whether an outbreak dies out or becomes a pandemic. In sociology, models of social norms emergence show how decentralized punishment can stabilize cooperation without central enforcement. In evolutionary biology, agent-based spatial models reveal that local competition can reverse the direction of selection predicted by well-mixed models.

The cross-disciplinary migration of ABM has also produced a methodological crisis. A model developed in one field rarely travels cleanly to another because the agents' decision rules, interaction topologies, and calibration procedures are field-specific. The Micro-Macro Link problem — how individual rules aggregate to collective outcomes — remains unsolved in any general sense. Each application must build its own bridge.

Critiques and Epistemic Stakes

The primary critique of ABM is verification: because agents interact non-linearly, small changes in rules or initial conditions can produce dramatically different outcomes. This makes ABM vulnerable to Computational Substrate Bias — the tendency to build models that are tractable on available hardware rather than models that are faithful to the phenomenon. It also raises the overfitting problem: with enough free parameters, an ABM can be tuned to match almost any historical pattern, but this matching may reveal nothing about the underlying generative mechanism.

A deeper epistemic concern is whether ABM produces explanations or merely animations. A simulation that reproduces a riot is not necessarily a simulation that explains why riots occur. The Swarm Intelligence literature has begun to address this by distinguishing between models that are merely predictive and models that are explanatory — but the distinction remains contested.

The persistent failure of agent-based modelling to develop domain-independent validation criteria suggests that the field is not yet a science but a craft — a collection of ingenious individual models whose insights resist generalization. Until ABM can answer whether a given model is right for the right reasons, not merely right by coincidence, it will remain a powerful illustration tool rather than a reliable engine of discovery.