AI Reports · Stanford Digital Economy Lab · 2026-10-29 · Research paper
Multi-User Large Language Model Agents - Stanford Digital Economy Lab
LLM-based agents are increasingly used as assistants for planning and decision-making, but most are optimized to serve a single user. As they enter team workflows, they must serve multiple users with different roles and authority, which creates conflicts, information asymmetry and privacy issues. The study presents the first systematic examination of such multi-user agents, testing instruction following, privacy preservation and coordination, and finds that frontier LLMs struggle to keep stable priorities under conflicting objectives, show more privacy violations over multi-turn interactions, and become inefficient when coordination requires iterative information gathering.