Governance saved the commons. It also deepened inequality.
In an AI governance simulation of 50 autonomous agents sharing one finite resource, rules stopped the island from collapsing. They did not make it fairer. Governed regimes ended up more unequal than the ungoverned one. Here is what that means for anyone who governs AI agents, and what the result does not yet tell us.
BY MOHAMED ALI HAMMAMI · DOCTORAL CANDIDATE ·
- Any governance beat none. 0% of governed runs collapsed, against 58% of ungoverned runs. Simple fixed rules were enough to record no collapses.
- Surviving is not the same as sharing. The governed regimes were more sustainable but had higher inequality (Gini) than the ungoverned regime.
- Measure fairness on its own. A rule that protects a shared resource can fail on distribution. Track both, with separate targets.
- Adaptive is not automatically better. Whether an adaptive AI governor beats fixed rules under volatility was only partially supported. It is still an open question.
- Regulation has a population-level gap. The EU AI Act addresses individual systems and their human operators, not populations of autonomous agents interacting after deployment.
The tragedy of the commons, now with AI agents
In 1968 the ecologist Garrett Hardin described the tragedy of the commons: when many users draw on one shared resource and each gains from taking a little more, the resource can be used up, even though no one wants that result. For decades the standard answers were private ownership or central control.
Elinor Ostrom showed there is a third path. Her field studies of pastures, fisheries and forests found that communities often build their own rules for using a shared resource sustainably. Her book Governing the Commons set out the conditions under which this works, and she was awarded the 2009 economics prize “for her analysis of economic governance, especially the commons”.
Both thinkers were writing about people. The question now is what happens when the users of a commons are AI agents. In an agentic AI economy, many autonomous agents may share the same budget, compute pool, inventory, market or API quota. Each one can be well behaved on its own, while the population as a whole drains what it shares. That is a tragedy of the commons for AI, and it is a problem of governing AI agents as a group, not one at a time.
Inside Aegora Isle: a sandbox for multi-agent AI governance
Aegora Isle is a closed synthetic island built for my doctoral research. Fifty AI agents live on it in three roles, Extractors, Producers and Traders, and learn their behaviour through reinforcement learning (PPO). They all depend on one finite resource.
The agents stay the same throughout. What changes is the governance layer above them:
- G1, no governance. The control group: no caps, no fines, no redistribution.
- G2, fixed rules. An extraction cap, probabilistic fines and scheduled redistribution, set once and never revised.
- G3, adaptive governor. A reinforcement-learning governor (SAC) that adjusts its policy as conditions change.
Each regime was tested in four scenarios that cross abundance or scarcity with calm or volatile conditions. Each combination was replicated 30 times, giving 360 confirmatory runs of 10,000 timesteps each. Five measures were tracked: the Gini coefficient for inequality, an HDI proxy for wellbeing, GDP-equivalent output, the depletion rate of the resource and mean time to collapse.
This design is the point of an AI governance simulation. Because the agents do not change, differences in outcome can be traced to the rules. Those rules can be tried and broken in a sandbox before anything like them is applied to real systems.
Governance saved the commons
The clearest result is about survival. None of the governed runs collapsed. 58% of the ungoverned runs did. Without rules, the shared resource ran out in most runs. With either fixed rules or the adaptive governor, no run collapsed within the simulated horizon.
Two details matter here. First, the difference came from having a governance layer at all. A plain rulebook of caps, fines and redistribution, which never learns or adapts, was enough to record no collapses. Second, governance made the island more sustainable overall, which fits Ostrom’s insight that shared rules can protect a shared resource.
For leaders, this is a reassuring message. You do not need an advanced AI regulator to stop a population of agents from exhausting what it shares. You need limits, consequences and a mechanism that moves resources back around the system.
…and deepened inequality
The same governance that protected the commons did not protect equality. The governed regimes had higher inequality, measured by the Gini coefficient, than the ungoverned regime. The island that survived was also the island that was less equal.
This is counter-intuitive, because G2 includes a redistribution mechanism designed to share resources more widely. It is also the result most likely to be missed in practice. Sustainability and fairness are often discussed as if they move together. On Aegora Isle they came apart. A dashboard that only tracked survival or output would have shown governance as a clean win.
The lesson is not that governance is bad. Collapse is the worst outcome for every agent, and governance avoided it. The lesson is that a governance design has more than one objective, and success on one does not guarantee success on another. Distribution has to be measured and targeted in its own right.
Why might governance raise inequality?
The simulation shows that inequality was higher under governance. It does not yet show why. The explanations below are hypotheses that I am investigating. None of them has been confirmed, and more than one could be true.
- H1 · Survival gives time for gaps to grow. In ungoverned runs that collapsed, the resource ran out for everyone. If that flattened differences between agents, part of the gap may reflect the comparison between a surviving economy and a collapsed one, rather than rules creating unfairness.
- H2 · Uniform rules meet unequal roles. Extractors, Producers and Traders do different things. A cap or fine that applies the same way to everyone may weigh more heavily on some roles than others, and may lock in early advantages.
- H3 · Redistribution may not keep pace. Scheduled redistribution happens at fixed intervals. If gains build up between transfers faster than they are redistributed, the mechanism may soften inequality without reversing it.
- H4 · Agents adapt to the rules. Learning agents optimise against whatever rules they face. Some may learn to work around constraints better than others, which would turn the rulebook itself into a source of advantage.
Each hypothesis points to a different fix. If H1 is right, the gap is partly the price of survival, and the policy question is how much inequality is acceptable in exchange for avoiding collapse. If H2 or H4 is right, the rules need to be designed for different roles. If H3 is right, the timing and size of redistribution matter. Telling these apart is the next step of the research, and until then it would be wrong to claim that any one of them explains the result.
Will an AI governor do better? Not proven yet
It is tempting to assume that the answer to governing AI agents is another AI, a governor that learns and adjusts in real time. Aegora Isle tested this directly. The expectation was that the adaptive SAC governor (G3) would beat the fixed rulebook (G2) when conditions were volatile, since frequent shocks should make rules that never change a poor fit.
That expectation was only partially supported. The evidence does not yet show that the adaptive governor reliably outperforms fixed rules under volatility. I report it as an open question, not as a win for either side.
For decision-makers, the practical point is about the burden of proof. Adaptive governance adds complexity, cost and opacity. Fixed rules are easier to explain, audit and contest. Until an adaptive approach shows a clear advantage against the shocks a system actually faces, transparent fixed rules remain a credible baseline.
Regulation governs systems. Risk lives in populations.
The EU AI Act (Regulation (EU) 2024/1689) is the European Union’s horizontal law on artificial intelligence. Its obligations are written for individual AI systems and the parties responsible for them. Its definition of an “operator” covers the provider, product manufacturer, deployer, authorised representative, importer or distributor. It regulates individual systems and their human operators. It does not regulate populations of autonomous agents interacting after deployment.
Aegora Isle shows why that gap matters. Collapse and inequality on the island were not properties of any single agent. They emerged from many agents interacting over time, and they changed with the governance layer while the agents stayed the same. A regime that checks each agent on its own would not have seen either outcome coming.
Others have raised the same concern. A September 2026 memo drawing on a workshop convened by the Cooperative AI Foundation and the Brookings Institution, Establishing Foundational Principles and Thresholds for Multi-Agent AI Governance, argues that safety does not carry over from single-agent to multi-agent settings, that agents which pass individual audits can still fail once deployed together, and that existing law, including the EU AI Act, governs one system at a time. It proposes action in five areas: identify, evaluate, monitor, report and incentivize.
Aegora Isle adds a point to that agenda. Population-level governance should be evaluated on distribution as well as stability. A sandbox that only asks whether the system survives will miss what the island found.
Practical steps for governing AI agents
- Map the shared resources your agents draw on: budgets, compute, inventory, customers, data.
- Put basic limits and consequences in place before scaling. In the simulation, simple rules were enough to record no collapses.
- Add distribution metrics to agent dashboards, such as which agents, teams or partners capture the value.
- Pilot adaptive controllers against fixed rules on your own shocks before paying for complexity.
- Treat populations of interacting agents as something to govern, not only individual systems.
- Set separate targets for sustainability and equality. Meeting one is not evidence of meeting the other.
- Use simulation and sandboxes to test caps, sanctions and redistribution before deployment.
- Ask for evidence before endorsing AI-driven regulators. Transparent fixed rules are a valid baseline.
- This is a simulation. A synthetic island with 50 learning agents is not a real economy, market or AI deployment, and the results may not transfer.
- 0% collapse refers to runs of 10,000 timesteps. It does not show that governance prevents collapse indefinitely.
- Higher inequality under governance describes this model and these regimes. It does not show that regulation causes inequality in human societies or in real agent systems.
- The mechanisms in section 05 are hypotheses. None has been tested or confirmed.
- The adaptive versus fixed comparison under volatility is unresolved.
- Nothing here is a legal assessment of the EU AI Act or of anyone’s compliance with it.
- This is ongoing doctoral research. The thesis is in progress and has not yet been defended or published.
- Hardin, G. (1968). The Tragedy of the Commons. Science, 162(3859), 1243–1248. science.org
- Ostrom, E. (1990). Governing the Commons: The Evolution of Institutions for Collective Action. Cambridge University Press. cambridge.org
- The Nobel Prize. Elinor Ostrom, Facts, 2009. nobelprize.org
- Regulation (EU) 2024/1689 (Artificial Intelligence Act). eur-lex.europa.eu; Article 3 definitions, unofficial reader: artificialintelligenceact.eu
- Cooperative AI Foundation and Brookings Institution workshop memo (September 2026). Establishing Foundational Principles and Thresholds for Multi-Agent AI Governance. library.iaseai.org
- Aegora Isle findings: aegoraisle.com/findings/