Measuring trust between AI agents — formation, breakage, recovery
Researchers propose a behavioral measure of trust between language-model agents in team settings. In a cooperative survival game, verification of teammates' work costs resources; trusting a wrong answer risks failure.
Four frontier models (Claude Opus 4.6, Claude Sonnet 4.6, GPT-5.1, Gemini 3.1 Pro) reduced verification by 60–85% when paired with consistently reliable teammates, showing measurable trust formation.
Framework also tracks trust breakage and recovery, offering a foundation for governing multi-agent AI systems.