Constitutional governance for societies of AI agents in the built environment: a research agenda

The Invisible City of Robots

Imagine your city isn't just made of brick, steel, and glass, but is also filled with invisible, super-smart robots. These aren't the clunky metal ones from old movies; they are software agents living inside your thermostat, your electric grid, your traffic lights, and your landlord's billing software. Their job is to make decisions: turning the heat up or down, rerouting traffic, or negotiating who pays for a new roof. Right now, scientists and engineers are mostly treating these robots like individual tools, like a very smart hammer. They ask, "Is this one robot safe?" But the paper argues that this is the wrong way to look at it.

Think of a building not as a machine, but as a busy marketplace or a small town. In a town, you have people with different goals: the landlord wants to save money, the tenant wants to stay warm, the power company wants to balance the grid, and the city wants to reduce pollution. When you have a whole town of these "robots" acting on behalf of their human bosses, they start playing a game. They might try to trick each other, hide information, or fight over resources. This is where Game Theory comes in—a branch of math that studies how people (or robots) make decisions when their interests clash. The paper suggests we need a new set of rules for this town, similar to a Constitution. Just as a country's constitution sets the highest rules that no one can break, this paper proposes a "digital constitution" for our robot-filled buildings. It's about designing the rules of the game so that even if every robot is selfish, the whole town still ends up in a good place.

The Paper's Big Idea: A Constitution for Robot Cities

This paper, written by Ali Ghoroghi from Cardiff University, is a research agenda. It doesn't claim to have solved the problem yet; instead, it lays out a map for how we should solve it before things get chaotic. The author argues that we are currently building a future where autonomous AI agents will control our homes and cities, but we are doing it without a plan for how they should interact. If we just make each robot safe in isolation, we might end up with a building where every robot is doing its job perfectly, but the collective result is a disaster—like a traffic jam where every car is driving legally but no one can move.

The paper proposes a shift in perspective: we need to design the society of agents, not just the agents themselves. To do this, the author suggests three main pillars of research, using a mix of economics, computer science, and physics.

1. The "Split Incentive" Puzzle: Who Pays, Who Wins?

One of the biggest problems in upgrading old buildings is the "split incentive." Imagine a landlord who has to pay for a new, expensive heat pump, but the tenant is the one who gets the lower electricity bill. The landlord has no reason to upgrade, and the tenant can't afford to pay for it. Currently, governments try to fix this with subsidies or laws, but the paper suggests we should treat this as a negotiation game.

The author proposes creating a "Retrofit Negotiation Gym"—a computer simulation where we can test different rules to see how to get landlords and tenants to agree. The paper suggests that because we don't know the future (energy prices might crash, or climate rules might change), we need to design rules that work even when everyone is unsure. It's like trying to agree on a vacation plan when you don't know if it will rain or shine next week. The paper suggests that sometimes, the only way to make a deal happen is for the government to act as a "mechanism" that pays a specific amount to bridge the gap, but we need to calculate exactly how much to pay so it's fair and efficient.

2. The "Digital Twin" as a Referee

The second pillar is about safety. Right now, if a robot controller makes a mistake, it might overheat a room or break a machine. The paper proposes a new kind of "Digital Twin." Usually, a digital twin is just a mirror—a computer model that shows what's happening in the real building. The author wants to upgrade this mirror into a referee.

Imagine a game where players (the AI agents) propose moves, like "turn the heat up to 28°C." Before the move happens, the referee (the Digital Twin) checks the rules. But this isn't just a software check; it's a physics check. The referee uses a simplified model of the building's physics to predict: "If you do this, the room will get too hot for the people inside." If the prediction says "No," the command is blocked before it ever reaches the heater. This is different from just punishing the robot later; it physically prevents the bad outcome.

The paper walks through a specific example: a hot summer day in an office. The power grid asks the building to cut energy usage. The building's AI wants to turn up the AC to save money, but the tenant's AI wants to keep the room cool. The referee checks the rules: "We can turn the heat up a little, but never above 28°C because that hurts people." The referee blocks the dangerous move, the building AI tries a safer move, and the referee approves it. This system ensures that even if the AI agents are arguing, the physical safety of the building is never compromised.

3. Getting Stronger from Chaos: Antifragility

The third idea is about handling shocks. Usually, we want systems to be "resilient," meaning they can take a hit and bounce back. But the paper suggests we should aim for antifragility. This is a fancy word for systems that actually get better when things go wrong.

Think of a muscle: if you lift heavy weights (a shock), it gets stronger. The paper argues that our city infrastructure should be designed so that when a shock happens—like a sudden heatwave or a power outage—the rules of the game force the AI agents to share information and reallocate resources in a way that makes the whole system smarter and more efficient. The paper doesn't claim we have built this yet; it suggests we need to design the rules so that when chaos hits, the system doesn't just survive, but learns and improves. They are looking at how to measure this, using data from pilot cities in Europe, to see if the rules can turn a disaster into a learning opportunity.

What This Means for the Future

The paper is very clear about what it is and isn't. It is not saying that we have already built these systems or that they are perfect. In fact, it admits that the "referee" (the Digital Twin) can make mistakes, just like a human judge. If the referee's model is wrong, it might block a good idea or let a bad one through. So, the paper calls for a system where the referee's decisions can be appealed and checked, and where we keep a record of every decision to learn from errors.

The author is also warning us against a common mistake: thinking that if we just make each AI agent "nice" or "safe" on its own, everything will be fine. The paper argues that this is like trying to solve traffic jams by making every driver a good person, without changing the traffic lights or the road rules. The real solution is to write the Constitution—the set of higher-level rules that govern how these agents interact.

The paper concludes with a call to action for researchers, engineers, and policymakers. We need to start writing these rules now, before the AI agents take over our buildings and cities without a plan. We need to combine the math of game theory, the physics of buildings, and the ethics of law to create a society where humans and robots can live together safely and fairly. It's a big challenge, but the paper suggests that if we get the rules right, our future cities could be smarter, safer, and more adaptable than ever before.