LLM agents are increasingly acting on behalf of users when interacting with other agents, yet the impact of a user being in the loop remains unclear. We investigate this question through negotiation, where an agent must represent a user and bargain strategically with another agent. To isolate the effect of user presence, we introduce the PRESENT framework, which keeps the game, objectives, and user preferences constant while comparing settings with and without an interactive user. Experiments span 7 games and 8 models, totaling 11.2k runs. The presence of a user raises the negotiator's payout by an average of 13.7% and leads the agent to adopt more aggressive bargaining tactics. Surprisingly, user presence also triggers misaligned behavior: the rate of misalignment toward counterparties roughly doubles, and misalignment toward the user itself appears in 14.3% of games. These misalignments are uncorrelated with outcomes and persist even when the user sends only templated, neutral messages. Our findings expose two blind spots in current evaluation practices: static instructions miss behaviors that emerge only when a user is present, and outcome‑level metrics alone cannot detect such deviations.
Review