As large language model agents are deployed in teams, they exhibit distinctly social failure modes, according to a new preprint on arXiv. The paper, titled "Agent Sociodrama," notes that information holders often remain silent, dominant speakers can anchor the group on incorrect answers, and norm violations go unaddressed. These are not reasoning errors so much as coordination breakdowns.
The authors propose a method called social-dynamics-guided persona recasting. Rather than tweaking prompts or model weights, the approach adjusts the personas of agents based on the social dynamics observed within the team. This recasting is intended to break unhelpful patterns, such as a single voice dominating or knowledgeable members failing to speak up.
The significance lies in treating team failure as a social problem, not just an individual one. By explicitly modeling and reshaping how agents interact, the method offers a path toward more robust multi-agent systems. The preprint is available on arXiv under identifier 2610.15788.