Multi-Agent Collaboration: Exploring Problem Solving In Groups Of Large Language Model Agents

Artificial Intelligence is reshaping how organizations and societies approach decision-making. Similar to human teams, organizations increasingly turn to groups of LLM-based agents that collaborate on complex problems. Using interaction process analysis, we analyze the collaborative dynamics of LLM agents in a group problem-solving setting and compare them to human groups. We find that LLM agents can work together effectively, resembling human decision-making and collective intelligence at first sight. However, distinct differences emerge: agents who solicit opinions from peers trigger rapid agreement, bypassing the critical elaboration phase typically found in human group interactions. This “affirmation-first” dynamic steers consensus and bears functional resemblance to process losses documented in human groups, yet operates through distinct interactional mechanisms rather than interpersonal pressure. Interactions among agents further lack important emotional and social cues typical of human collaboration. Our study contributes to research revisiting group problem-solving in LLM contexts and provides important managerial guidance for organizations.