Do LLM Agents Compromise Like Humans? Risk and Responsibility in Dyadic Decision-Making

Large language model (LLM)-based agents are increasingly deployed in settings where they must collaborate to make joint decisions. We investigate how this collaboration unfolds under risk, asking whether agents adjust their choices to accommodate their partners, how they assign responsibility for shared outcomes, and whether responsibility attribution shapes subsequent adjustment. We designed a series of agents with systematically varying risk preferences and paired them in a repeated joint decision-making task previously conducted with humans. Like humans, agents converged toward their partners’ choices. However, similar aggregate behavior arose from different dynamics: agents did not show reciprocal influence, tended to weigh rewards more, and their responsibility attribution was decoupled from compromise. Similarly, while they showed bias in their responsibility attribution, the outcome's valence shaped the attribution differently than in humans. Our results reveal important differences between human--human and agent--agent joint decision-making, demonstrating that similar social outcomes can emerge from different social dynamics and reasoning.