LLM Agents Cooperate via Self-Negotiated Contracts

A new arXiv preprint (2607.22750) explores how AI agents can use contract-based mechanisms to ensure cooperation in multi-agent settings. Drawing on legal institutions, the researchers study LLM-based agents in a spatial-temporal game combining bargaining and navigation. They compare formal contracts that compile to code with natural language contracts, finding that contract representations enable credible commitments and reduce defection incentives. The work addresses the principal-agent problem where cooperation costs are immediate but benefits are delayed.