Characterizing Behavioral Dynamics of Social Agents in Virtual Worlds

AI agents increasingly operate not only individually but as teams and within larger networks of interaction, yet whether persona-driven behavioral patterns remain consistent across these scales is largely untested. We construct a diverse simulated virtual world of persona-conditioned agents instantiated using Microsoft's TinyTroupe framework and evaluate behavior across three levels: individual decision-making using standardized behavioral methods, collaborative dynamics in multi-agent resource-allocation scenarios, and social network formation through interaction among agents. Our analysis showed that agents' behavioral tendencies were directionally consistent across all three levels. Individual-level cross-task consistency patterns extended to collaborative and network contexts, consistent with patterns observed in human social networks. These findings suggest that standardized human behavior frameworks can characterize directionally consistent, persona-driven emergent behavioral patterns in AI agents across individual, collaborative, and network scales.