Experimental evidence on the emergent social network patterns of LLM-driven multi-agent system for edge service collaboration

As a common phenomenon that stable social structures emerge in real societies, an important question is whether similar non-random collaboration patterns can arise in large language model (LLM)-driven multi-agent system deployed in intelligent edge environments without explicit social rules. Therefore, we develop a round-based multi-agent simulation platform in which LLM-driven agents engage in long-term interactions. The interaction process produces an evolving directed social network for subsequent structural analysis. Specifically, we analyze the emergent structures based on over 200,000 simulation records. At the macro level, internal relationship strength decreases with increasing team size, indicating the emergence of a diseconomies-of-scale effect analogous to that observed in real organizations. From the micro perspective, agents are more likely to form new relationships when sharing common interaction partners, ex-hibiting a pronounced triadic closure effect relative to random baselines. Furthermore, robust reciprocal interaction patterns emerge with balanced dependence between agents promoting the formation of stable bidirectional relationships, consis-tent with power-dependence theory. Finally, these results demonstrate that even in the absence of predefined social mechanisms, LLM-driven multi-agent system operating in edge-inspired distributed settings can spontaneously generate meaningful network structures, supporting their use in the design of scalable and cooperative LLM agent society.