Multi-agent AI governance: what the island has shown so far
Aegora Isle is an AI governance simulation built for doctoral research. Fifty autonomous AI agents share one finite resource under no rules, fixed rules or an adaptive reinforcement-learning governor. This page sets out what 360 confirmatory runs found, what each finding means for executives and policymakers, and what the results do not show.
A tragedy of the commons, run by AI agents
The tragedy of the commons describes what happens when many actors draw on one shared resource and each has a reason to take a little more. Aegora Isle recreates that problem with AI agents instead of people. Fifty agents, in the roles of Extractors, Producers and Traders, learn their behaviour with reinforcement learning (PPO) on a closed synthetic island. All of them depend on the same finite resource.
The agents stay the same across the experiment. Only the governance layer changes:
- 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 on the island change.
Each regime ran in four scenarios that cross abundance or scarcity with calm or volatile conditions, with 30 replications per combination. That gives 360 confirmatory runs of 10,000 timesteps each. Five outcomes were tracked: the Gini coefficient (inequality), an HDI proxy (wellbeing), GDP-equivalent (output), depletion rate and mean time to collapse.
Governance made the island more sustainable, but more unequal
The governed regimes made the island more sustainable than the ungoverned baseline. They also left it more unequal: governed regimes had higher inequality than the ungoverned regime, as measured by the Gini coefficient.
This matters because sustainability and fairness are often assumed to move together. On Aegora Isle they did not. Rules that protected the commons did not, on their own, protect equality among the agents that depend on it.
Read the in-depth analysis: governance saved the commons but deepened inequality →
If many agents draw on a shared budget, inventory or compute pool, guardrails that prevent depletion may not spread value evenly. Track who gains, not only whether the system stays up.
Sustainability and equality need separate targets and separate measurement. A rule that succeeds on one can fail on the other.
Governance strongly delayed collapse
This is the clearest result. 0% of governed runs collapsed, compared with 58% of ungoverned runs. Without governance, the shared resource ran out in a majority of runs. With either fixed rules or the adaptive governor, no run collapsed within the simulated horizon.
The biggest difference in survival came from having a governance layer at all. Even a simple, fixed rulebook changed the outcome compared with letting agents act unchecked.
The case for governing agent populations does not depend on sophisticated AI regulators. In this simulation, basic caps, fines and redistribution recorded no collapses.
Adaptive AI governance vs fixed rules under volatility: still an open question
The study tested whether an adaptive governor earns an advantage in volatile conditions, where frequent shocks should make fixed rules a poor fit. That hypothesis was only partially supported. The evidence does not yet show that the adaptive reinforcement-learning governor reliably outperforms the fixed rulebook when conditions are volatile.
It is reported as an open question, not as a win for either approach.
Do not assume an adaptive, AI-driven controller will beat clear fixed rules. Test both against the shocks your own systems face before paying for complexity.
Adaptive AI regulators are a hypothesis to evaluate, not a settled solution. Transparent fixed rules remain a credible baseline.
The EU AI Act and autonomous agents: a population-level gap
The EU AI Act regulates individual AI 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 themselves stayed the same.
Compliance for each agent may not cover risk at the level of the whole system. Plan to monitor how your agents behave together, not only one by one.
Consider rules that apply to interacting populations of agents, such as shared caps, sanctions and redistribution, and use simulation to test them before deployment.
- It is a simulation. A synthetic island with 50 learning agents is not a real economy, market or AI deployment.
- 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. It does not show that regulation causes inequality in human societies.
- It does not settle whether adaptive AI governance is better than fixed rules. That comparison remains open.
- It is not a legal assessment of compliance with the EU AI Act.
- This is ongoing doctoral research. The thesis is in progress and has not yet been defended or published.
Explore the simulation yourself
Researchers, students and policy practitioners can request access to Aegora Isle. Registrations are reviewed before access is granted.