Can AI automatically redesign itself to solve cooperation problems?

Getting multiple AI agents to cooperate in social dilemmas (like resource-sharing games) is hard. Researchers built a two-level system where an outer AI agent acts as a researcher, automatically editing the code, prompts, and logic of an inner LLM system that synthesizes cooperative policies. Across two games and two welfare objectives, the self-redesigning pipeline beat hand-tuned baselines and found that fairness mechanisms emerge only when optimizing for maximin (protecting the worst-off), not utilitarian efficiency. Code released.