Why Simulated Worlds Need Better Economists

Jul 20, 2026

9 min read

4 views

78%

The hidden problem with learning machines

What if the hardest part of building intelligent agents is not teaching them to act, but teaching them to live inside a world where their actions change the world itself? That is the strange overlap between multiagent simulation and the economics of digitization. On one side, we build environments where agents learn by interacting with each other. On the other, we study digital markets where behavior, data, and platform structure constantly reshape the conditions of competition. In both cases, the system is not a static puzzle. It is a living ecology.

That matters because a lot of modern AI thinking still assumes a simple loop: observe, learn, act, improve. But once multiple agents are involved, the loop becomes reflexive. One agent’s policy alters another agent’s incentives. One platform’s data collection changes the market. One training episode changes the next episode’s distribution. The central question is no longer just, “What is the best policy?” It becomes, “What kind of world is the policy making?”

This is where simulation and digitization meet. Agent-based environments give us a controlled laboratory for interactions, while digital economies show what happens when interactions scale into markets with feedback, externalities, and power. Together they reveal a deeper truth: intelligence is not only about optimizing within a world, but about understanding when the world itself is endogenously changing.

The trap of treating data like fuel

In conventional machine learning, data is often treated like a raw material. More data should mean better predictions, and better predictions should mean better decisions. That intuition is useful, but it hides a dangerous simplification. Data in digital systems does not behave like oil or metal or even electricity. It is not a passive input waiting to be consumed. It is often a byproduct of interaction, and its value depends on what the system is already doing.

That is why the idea of a simple data feedback loop is so seductive and so misleading. If more usage generates more data, and more data generates a better product, then the winner seems destined to pull away forever. Yet the deeper issue is that learning from data often shows diminishing returns. The first thousand observations may be hugely informative, but the next million may add only marginal improvement. In some settings, the market does not become infinitely more concentrated just because one actor has more data. The mechanism is subtler.

Think about a restaurant recommendation app. At first, user reviews are powerful because they reveal hidden quality differences. But once the app has enough reviews, the marginal gain from another generic five star rating may be tiny. What matters more is the structure of the experience, the timing of demand, the neighborhood, the weather, and the preferences of the current user. Data helps, but it does not automatically create an unbeatable fortress.

This is an important correction. It suggests that many digital advantages come not from data as a conventional input, but from data embedded in a system of coordination. The platform that sees more behavior can often shape behavior, not just predict it. That is a different source of power entirely.

The real advantage is not information, it is control over the game

A useful mental model here is to distinguish between learning advantage and game design advantage. Learning advantage means you can estimate the world better than your rivals. Game design advantage means you can alter the rules, friction, visibility, and incentives that structure the world in the first place.

This is why some digital platforms become powerful without relying on a straightforward data monopoly. They can sort, rank, recommend, throttle, bundle, or standardize in ways that change the behavior of everyone else in the system. The value is not only that they know what users want. It is that they become the medium through which users discover, compare, and transact. In other words, they stop merely observing the market and start hosting it.

Agent-based simulation is ideal for studying this kind of power because it makes interaction explicit. In a static model, an agent’s strategy can be evaluated against fixed payoffs. In a multiagent environment, however, the payoffs emerge from the collision of policies. A trader’s rule changes the price path. A driver’s decision changes congestion. A buyer’s choice changes what the seller learns. The world answers back.

This is the core insight: in interactive systems, information is valuable not only because it reduces uncertainty, but because it changes strategic position. If one agent can anticipate another, it may gain a small edge. If one agent can structure the environment that the other inhabits, it may gain a structural edge. The difference is enormous.

The deepest form of digital power is not knowing more than everyone else. It is arranging the stage on which everyone else must act.

That helps explain why some data advantages fade while others persist. Predictive power can saturate. Structural power can compound.

Why simulations are the laboratory for digital economics

Agent-based simulation is not just a tool for training intelligent agents. It is a way to experiment with the logic of economies before those economies harden in the real world. When environments include multiple agents, externalities, strategic adaptation, and platform-mediated behavior, analytical shortcuts often fail. Simulation becomes a microscope for feedback loops that are too messy to derive cleanly on paper.

Imagine a ride-sharing market. If you only model supply and demand in aggregate, you might miss the fact that pricing changes driver availability, driver availability changes wait times, wait times change rider expectations, and rider expectations change future demand. Add data collection, and the platform learns where bottlenecks occur. Add multiple competing platforms, and each platform’s actions alter the others’ markets. This is not a single optimization problem. It is a recursive system of adaptation.

Agent-based environments capture that recursion naturally. They let us ask what happens when learners do not just respond to the environment, but create it. That is why interfaces such as gym.Env, MultiAgentEnv, and OpenSpiel matter. They are not merely software conveniences. They are invitations to model worlds where strategy, observation, and adaptation are inseparable.

The economic insight here is subtle but powerful. In digital systems, the market is not a backdrop for behavior. It is an evolving product of behavior. Traditional analysis often treats demand as given and then asks how firms compete inside it. Multiagent simulation reminds us that demand itself may be endogenous, especially when platforms mediate attention, trust, discovery, and feedback.

That makes simulation a kind of institutional imagination. It lets us test not just what a policy does to a population, but what a policy does to the rules by which the population evolves.

From prediction to ecology: a better framework for thinking about digital systems

To connect these ideas more concretely, it helps to use a three layer framework:

Prediction layer: Who knows more? This is the classic data question. More signals, better models, tighter forecasts.
Interaction layer: Who changes whom? Here the focus shifts to strategic response, congestion, imitation, competition, and network effects.
Institution layer: Who sets the rules? This includes platform design, ranking systems, API constraints, access policies, and market structure.

Most debates about data power stay stuck at the prediction layer. They ask whether more data gives a firm better estimates and thus more market power. But much of the real action sits in the interaction and institution layers. A platform that slightly improves prediction may gain some advantage. A platform that rewrites attention pathways or transaction rules can transform the whole ecosystem.

This framework also clarifies why diminishing returns to data do not settle the question of market power. Even if predictive improvements flatten, the system may still produce concentration through control over interactions or institutions. A firm may not need perfect knowledge to dominate a market if it can determine what counts as a recommendation, what gets surfaced, or what friction users face when switching.

Consider Airbnb. The key issue is not simply that the platform has more booking data than any individual host. It is that it sits at the junction of search, trust, pricing, and demand routing. It can therefore shape which listings are visible, how they are compared, and how market participants respond. The advantage lies in orchestration as much as in learning.

This is why economists and AI practitioners should speak more often. Economists are trained to ask where power comes from. AI researchers are trained to ask how agents learn in environments. The combination is potent because digital systems are both markets and environments. They are arenas of competition and laboratories of adaptation.

What this means for builders, researchers, and policymakers

If digital systems are ecologies rather than static markets, then our habits of design and regulation must change.

For builders, the lesson is to stop overestimating raw data scale and start measuring structural leverage. Ask not only how much better the model gets with more data, but how the system changes when the model is deployed. Does it alter user behavior? Does it create lock-in? Does it amplify one side of a market? A small prediction gain can become a massive strategic advantage once embedded in a feedback loop.

For researchers, the lesson is methodological. Benchmarking agents against fixed environments can hide the most important phenomena. We need more experiments where agents adapt to each other, where the environment evolves, and where policy changes can be traced through second and third order effects. In practical terms, that means treating multiagent simulation not as a niche, but as a central method for studying digital economies.

For policymakers, the lesson is to focus less on whether a platform has “too much data” in the abstract and more on whether it has too much control over the conditions of competition. A platform may not need unlimited data to dominate. It may only need enough data to engineer the environment in which others must compete. That suggests antitrust, interoperability, and transparency questions should look beyond storage volume and toward market architecture.

A practical rule of thumb: when a system’s best strategy is to learn faster, you have a prediction problem. When its best strategy is to shape behavior, you have a governance problem.

Key Takeaways

Do not treat data like a conventional input. In digital systems, data is often a byproduct of interaction, not a neutral raw material.
Separate predictive advantage from structural advantage. Better forecasts matter, but control over rankings, routing, visibility, and rules matters more.
Model feedback loops explicitly. Multiagent simulation is useful because real digital markets are recursive, not static.
Watch for diminishing returns to data, but do not confuse that with the absence of power. Market power can arise from orchestration even when prediction gains flatten.
Ask who sets the environment. The most consequential actor is often the one that shapes the stage, not the one that knows the most about the performance.

The world does not just reward intelligence, it rewards world-making

The most interesting connection between agent-based simulation and the economics of digitization is that both force us to abandon the fantasy of a passive world. Agents do not merely observe reality. They participate in making it. Platforms do not merely record markets. They organize them. Data does not simply accumulate. It reverberates through incentives, institutions, and interactions.

That is why the old question, “Who has the most data?” is too small. The better question is, “Who has the ability to turn observations into an environment?” In digital systems, that is where durable advantage lives. And in multiagent worlds, that is where intelligence becomes strategy.

So perhaps the next frontier is not just better algorithms or bigger datasets. It is a richer understanding of agent ecologies, where prediction, behavior, and institutions co-evolve. Once you see that, simulations stop looking like toy worlds and markets stop looking like spreadsheets. They start looking like what they really are: living systems that learn back.