Output / Research Journal
Arguing in agent-based models
Note 10 from the research journal by Nicklas Lundblad:In the work here so far we have mostly spoken about agents as a kind of interface and capability bundling of artificial intelligence, but the research project is also looking at things like institutions and agency, and in doing so the focus is more on what is called “agent-based modelling”(ABM).
Note 10.
Such modeling allows us to explore how simple rules create different kinds of emergent behavior and it is among the most useful ways to learn about complexity and complex systems - yet, up to now, it is rather rare to find students in social sciences working with ABM. This is a pity - since there are actually excellent tools available to do so.
One of these tools is NetLogo . It is free and has a large library of different agent based scenarios, where students can explore emerging behaviors in different ways. Examples include things like Schelling segregation.
Thomas Schelling's segregation experiment is a simple but powerful model showing how small individual preferences can lead to large, unintended patterns in society. Schelling imagined a grid—like a checkerboard—where each space could hold an individual of one of two groups. Each person prefers to have at least some neighbors of their own group, but they do not require full segregation. For example, someone might be perfectly happy as long as at least 30% of their neighbors are similar to them. If this condition is not met, they move to another empty spot on the grid. Importantly, no one in this model starts out wanting complete separation; they simply have mild preferences for not being in the extreme minority.
When the model runs—people checking their surroundings and moving if they feel uncomfortable—an unexpected pattern emerges: even very modest individual preferences inevitably produce sharply segregated neighborhoods. The system “tips” into separation not because anyone intended it, but because the dynamics of many small decisions reinforce each other. Schelling's insight is that macro-level patterns (segregated cities) can arise from micro-level behavior (small, seemingly harmless personal preferences). His experiment shows how complex social structures can emerge from simple rules, and how societies may drift toward outcomes that no individual actually desires.
The advantage of NetLogo and other ABM environments is that it does not just tell us this, it shows us how the model plays out:
From a social sciences perspective this suggests a methodological choice: with the rise of ABM and advanced simulations, there is now a possibility to require that research is argued, formalized and developed in models. A requirement that we articulate our theories in models means that we need to think through the plausible mechanisms that are in play in society and how they relate – doing that means that the translation into policy also becomes easier, or at least that the difficulties in translating research into policy is made clearer.
NetLogo does not require that we learn coding (although we can, and probably should), but it allows for us to express social theories in models that can be debated robustly with an aim at policy progress.
Let's take an interesting example that brings us back to AI agents and institutions. One of the things we should expect if we enter a world in which agents interact with each-other on our behalf in commercial settings and contractual relationships, then they should find value in harmonizing all the frameworks around such contracting into a lex agentia modeled on the historical law of trade, lex mercatoria. Can we test this hypothesis in NetLogo? A bit of tinkering gives us a model that works like this (illustrated in Gemini here):
Agents begin in four distinct "national" clusters (represented by colors); When they trade with partners under the same legal system, they gain wealth, but interacting with conflicting laws causes financial loss due to high transaction costs. To stop losing money, struggling agents engage in "positive collusion" by spontaneously adopting the legal system of their wealthiest neighbor, acting on the premises so that the neighbor's system is more efficient. This creates a cascading network effect where agents rapidly abandon inefficient or minority jurisdictions to coordinate on a shared standard, eventually causing the entire population to converge on a single, unified legal system solely to minimize commercial friction.
We can also go deeper and use agent-based frameworks like MESA and build models in python. This is perhaps a bit more complex, but not much. Let's look at an example.
The model initializes a population of contracting entities (firms), each assigned to a home jurisdiction with a randomly-generated legal system quality score, plus several neutral international frameworks (UNIDROIT, ICC Rules, international arbitration) that have higher baseline quality. Each simulation step, agents pair up and negotiate which law will govern their contract—when collusion is enabled, they jointly optimize by scoring candidate legal systems on a weighted combination of network adoption (how many other agents already prefer that system) and intrinsic quality, minus any switching cost penalty; when collusion is disabled, the more connected agent's preference wins. After contracting, agents accumulate transaction costs (lower when both parties are familiar with the chosen law and when the legal system is high-quality), and they gradually shift their own preference toward whichever framework they've been using most frequently in recent contracts. Over time, this creates a feedback loop: early adoption advantages compound through network effects, regional clusters form among frequent trading partners, and eventually the system tips toward a dominant legal framework—typically converging to 90%+ adoption within 50-200 steps when collusion is allowed, versus slower and less complete harmonization under competitive negotiation. Here is an example of a run:
We can also study what the results are without any agential collaboration or collusion - and that allows us to see that even without active collusion we do end up with some harmonization over time, but the transaction costs in our model look wildly different:
Now, you may think this is wrong - but the beauty of arguing in models is that the only thing you have to do is tweak the parameters to show your point, or reject the model and propose your own.
And, the next step, of course, is to combine agents with LLMs - and rather than having the individual agents just follow simple rules (as in the Schelling case) we can allow for them to make decisions on the basis of a set of instructions that allow for greater variety. The perhaps most interesting and advanced project in this space is Google DeepMind's Concordia . The research team behind the project describe it well:
"Concordia is a library to facilitate construction and use of generative agent-based models to simulate interactions of agents in grounded physical, social, or digital space. It makes it easy and flexible to define environments using an interaction pattern borrowed from tabletop role-playing games in which a special agent called the Game Master (GM) is responsible for simulating the environment where player agents interact (like a narrator in an interactive story). Agents take actions by describing what they want to do in natural language. The GM then translates their actions into appropriate implementations. In a simulated physical world, the GM would check the physical plausibility of agent actions and describe their effects. In digital environments that simulate technologies such as apps and services, the GM may, based on agent input, handle necessary API calls to integrate with external tools.”
Now we can start to model social problems in entirely new ways, and our model is essentially the narrative of the GM!
The combination of agents, artificial intelligence and modeling arguably means that we are living at one of the most exciting moments in the history of social science - and the sooner we take it on, the more we will learn. A society that argues in models is likely to be much more healthy than one that gets stuck in polarization.
