The "One Brain, Many Mouths" Problem in Multi-User AI Agents