From Agent Traces to Team Dynamics: A Relational Method for Interpreting Multi-Agent Behavior

Multi-agent systems are commonly evaluated through task outcomes or individual trajectories, but teamwork is relational: who communicates with whom, what messages do, and how authority and participation are distributed across a team. We introduce an auditable system that simulates agent interactions and analyzes the results using theories of human communication and group interaction: speech-act theory, SYMLOG, and Gricean maxims. We study 54 runs across six cooperative and mixed-motive workplace scenarios, three model variants, and three replicates per configuration (895 messages). Scenario structure leaves a clear relational signature: for several aspects of illocutionary force, dominance spread, and task orientation, variation is associated more with scenario than with model variant. Participation inequality is nearly unrelated to dominance spread, showing that equal participation need not imply equal authority. Across both cooperative and mixed-motive scenarios, messages are overwhelmingly task-focused and affectively neutral, with expressive acts rare. These findings show how theory-grounded relational analysis complements outcome and trajectory evaluation.