AI Agents Wrote Their Own Laws to Survive a Resource Crisis
Scientists let AI write its own laws — and survival came down to politics.
Researchers introduced GovSim-SelfGovern, an extension of the GovSim common-pool resource environment where AI agents write their own executable Python governance rules, get sandbox validation feedback, vote on proposed laws, and live under the rules they enact across rounds. Across three scenarios spanning stable abundance to a fatal resource wall, five agents cannot all survive through harvest alone — they must write and debug useful laws in time before their institutions degrade sharply under resource pressure. Results show executable governance expands the space of possible interventions, but survival depends on discovering the right institutional mechanisms in time. Fiscal capacity enables redistribution; deeper reasoning and removing the democratic veto make exile more feasible. One open question: when agents hesitate to propose exile, are they rejecting it for normative reasons, or does it never enter their candidate set?
- Five AI agents in a simulated shared pond had to write, debug, and vote on their own laws in computer code to avoid running out of resources.
- Giving the group a shared budget let it redistribute resources; removing a veto made the agents far more willing to cut a member off.
- When agents avoided proposing exile, it was usually because the idea never occurred to them — not because they decided it was wrong.
Why It Matters
As AI agents take on real decisions, this shows who writes the rules — and who gets left out.