Agent-based models
Science as Grand Theft Auto
2011-07-28 — 2026-10-22
Wherein Bottom-Up Simulations Are Constructed From Individual Components, Ranging From Schelling’s Checkers-Based Segregation to EpiCast’s National Disease Modeling and LLM-driven Shachi Frameworks for Collective Dynamics.
…comprise a particular emphasis within the broad field of simulating things.
An agent-based simulation is designed to approximate the behaviour of a real-world system. It tends to avoid partial differential equations or fixed-point proofs in favour of breaking a large system down into its component chunks. If we were looking at a simulation of an economy, an agent-based version might start from the individual humans buying and selling within that economy. Within that simulation, we might want to work out, for example, “given what we know about humans and can simulate, what does that tell us about markets?” “Bottom-up” modelling would be a more suggestive name.
1 Examples
Roll your own, code-free, using Nicky Case’s Emoji prototype.
Howdy! 🌻 This (prototype) is a tool that lets you simulate systems of the world. In emoji.
My hope is that this can help people make more stuff like Parable of the Polygons, interactives that help people understand the complex systems of the world we live in. Also, emoji. In this simulation, you can edit EVERYTHING. Draw on the grid, modify the rules, even change these very words you’re reading! And at the end, you can save and share your creation. That’s the cool thing: there is no difference between playing & making, between reading & writing.
or Alex Komorske’s entry, jkomoros/CASsim: Complex Adaptive Systems Simulator
CASsim stands for Complex Adaptive Systems simulator. This is a very basic harness and framework to make configurable, interactive agent-based models and then generate animated screenshots of their results.
Nobel Laureate Thomas Schelling’s 1969 Racial segregation model based on checkers pieces.
As a kind of primordial ancestor, the classic Ising spin glass model
A modern example — EpiCast simulates the entire population of the USA, or at least, far more detail about them for disease control than I would have suspected to be reasonable.
-
what we do know about is the interaction between large crowds and the police and you could do much worse than check out the work of psychologist Clifford Stott who researches how crowds react to policing and what triggers violence.
In his 2009 report on the scientific evidence behind ’Crowd Psychology and Public Order Policing,’ commissioned by the UK constabulary, he summarises what we know about public disorder and how the authorities can best manage it (you can download it as a pdf).
He notes that the old ideas about the ’mob mentality’, deindividuation and the loss of individual responsibility are still popular, but completely unsupported by what we know about how crowds react.
2 What are they for?
The basic problem of science runs something like this — our current models of system \(X\) with parameter set \(P\) are good at approximating some of the observed behaviours of reality for some range of parameters \(p \subset P\), and that range can be fairly easy to find. However, our model does not seem able to produce observed behaviour \(b\) for any value of parameters. Let’s remove some of the simplifying assumptions in our model of \(X\) and make it a more complex model, \(X'\) with more parameters \(P'\), and more diverse behaviours. Now it can possibly produce behaviour \(b\) for some values of parameters \(p'\subset P'\). But we are no longer assured that we can find the set \(p'\) because of, e.g., loss of nice analytic forms for our model, or because of the curse of dimensionality.
Typically, agent-based models perform poorly here because they have vastly more parameters to fit, and no more data points to fit them to, than other models. In macroeconomics, for example, an economy full of different agents with unique behavioural heuristics, methods of production, and technologies has many more parameters than one where you assume an equilibrium is reached by aggregating demand and supply functions. Although you can add more parameters to models based on either set of techniques, or devise hybrid models.
These days we have a lot of compute and data — at least of some sorts — and better fitting algorithms. We can use these to, perhaps, discard some of our simplifying assumptions by producing them as an emergent property of an underlying process, so we have the luxury, perhaps, of using agent models whose behaviour is harder to simplify, and running the model anyway.
Alternatively, even if we can’t hope to fit the model to real data, we can still run models that test various possible sets of assumptions about behaviour in the hope of getting it at least qualitatively right, and congratulate ourselves that we have not only gained insight, but protected ourselves from the danger of mistaking our model of reality for reality itself.
Or we might like to use agent-based models because, in some sense, such models are more intuitive than more mathematical abstractions; everything cashes out in the choices made by individuals. Moreover, in this case, the intuitive models might give more direct insight into what is happening in a particular situation by letting non-specialists participate more directly. This collaborative option is called companion modelling and it dwells in the dangerous no-man’s-land between science, policy, and community engagement; I, for one, am terrified of that area. I’d much rather analyse a problem from the outside, be annoyed when no one takes my advice, and earn the same sense of righteous frustration, with much less time and money spent.
Things to be interested in:
- Agent-based models that include realistic constraints of agent attention and search strategies.
- Economic models that include not only innovation, but plausible representations of the size of innovation search space and the costs of exploring it. Not necessarily remotely possible.
Now, read:
- the Scholarpedia article on the ABMs
- Leigh Tesfatsion’s Agent-Based Computational Economics page
- Artem Kaznatcheev, who is quite a good mathematician, decides that agent-based models are awful but so are analytic ones.
3 Using AIs
SakanaAI/shachi: Reimagining Agent-based Modeling with Large Language Model Agents via Shachi (Kuroki et al. 2026)
How collective behaviors emerge from the interactions of individual LLM-driven agents is a central question in artificial life, yet controlled study of these emergent dynamics has been hindered by the lack of a principled simulation framework for systematic experimentation. To address this, we introduce Shachi, a principled methodology and modular framework that decomposes an agent’s cognition into core components: Configuration for intrinsic identity, Memory for contextual continuity, and Tools for extended capabilities, all orchestrated by an LLM reasoning engine. This decomposition treats each cognitive component as an independently controllable variable, enabling perturbation studies that trace how micro-level cognitive traits propagate into population-level dynamics. We investigate behavioral patterns across a 10-task benchmark spanning three levels of collective complexity. Shachi enables memory transfer across environment transitions, producing history-dependent behavioral shifts, and allows agents to simultaneously inhabit multiple environments, revealing cross-environment interference invisible in single-environment studies. Furthermore, in a real-world U.S. tariff shock case study, locally interacting agents with individually controlled cognitive components produce macro-level market dynamics directionally consistent with observed real-world outcomes. Our work provides a rigorous, open-source simulation framework for LLM-based ABM, aimed at fostering cumulative scientific inquiry into the emergent collective behaviors of interacting artificial agents.
4 Curiosities
CSIRO ABM used to be a thing.