Commitment To Cooperation With Self-Negotiated Contracts

This paper studies how AI agents can cooperate when the costs of cooperation arrive before the benefits, creating incentives to defect. Inspired by legal contracting, the authors evaluate self-negotiated contracts between LLM-based agents in CT, a spatial-temporal game combining bargaining with navigation toward a goal. The study compares contract representations ranging from formal agreements that compile into code to natural-language contracts requiring reinterpretation. Across multiple LLM backbones, model sizes, and providers, the authors report that self-negotiated contracts can produce better cooperative outcomes than regular trading alone.