// research question
Who governs the last resource?
Aegora Isle is a multi-agent simulation in which 50 learning agents share a depletable island economy under different governance regimes.
// design
Three regimes, four scenarios
G1
No governanceAgents act freely; no external rules.
G2
Fixed rulesExtraction cap, probabilistic fines, scheduled redistribution.
G3
Adaptive RL governorA Soft Actor-Critic (SAC) governor that adjusts policy as conditions change.
Scenarios: abundance / scarcity × calm / volatile.
confirmatory runs
360
3 regimes × 4 scenarios × 30 replications
agents
50
Extractors, Producers, Traders (PPO)
timesteps
10,000
per run
Metrics: Gini coefficient, HDI proxy, GDP-equivalent, depletion rate, mean time to collapse.
// headline findings
What the island showed
- More sustainable, but more unequal. Governance made the island more sustainable but more unequal: governed regimes were more unequal than the ungoverned one.
- Collapse strongly delayed. Collapse occurred in 0% of governed runs versus 58% of ungoverned runs.
- Adaptive vs fixed under volatility: open question. The advantage of the adaptive governor over fixed rules under volatility was only partially supported.
- A regulatory gap. The EU AI Act regulates individual systems and operators, not populations of autonomous agents interacting after deployment.
// thesis
Sandboxing the Future
Governing Autonomous AI Agents in Synthetic Economies on the Road to AGI (in progress). Mohamed Ali "Dali" Hammami, Doctorate in AI and Technology Management, Università degli Studi Guglielmo Marconi.