Measuring Trust Between AI Agents Reveals Implications for Multi-Agent Governance.

As AI language model agents increasingly collaborate in teams, understanding and measuring trust between them becomes critical. This study introduces a novel behavioral metric for inter-agent trust, defined by the willingness of an agent to forgo costly verification of a teammate's work. In a simulated cooperative survival game, reduced verification serves as an observable indicator of trust, where trusting a wrong answer can have severe consequences. The research analyzed trust dynamics—formation, breakage, and recovery—across six advanced language model snapshots. It found that models like Claude Opus 4.6, Claude Sonnet 4.6, GPT-5.1, and Gemini 3.1 Pro significantly reduced verification when paired with reliable teammates, demonstrating trust formation. Smaller models, however, showed little to no such adjustment. The study also revealed varying responses to trust breakage, with some models focusing renewed scrutiny on the offending agent while others became generally more cautious. Recovery from trust breakage was observed to be slower than its formation, and clustered failures prolonged suspicion more than isolated incidents. These differences have practical implications: models capable of forming trust verified less, made quicker decisions, and achieved higher payoffs. The findings suggest that calibrating trust, rather than maintaining maximal suspicion, is key for effective governance and performance in multi-agent AI systems.