Communication Causes LLM Agent Outputs to Converge

As Large Language Models (LLMs) are increasingly deployed as communicating agents, a critical question arises: does their interaction lead to a convergence of their outputs? This paper introduces BOUNDARY_SYNC, a novel protocol designed to quantify "representational coupling" using a metric called the Coupling Amplification Factor (CAF). A CAF less than 1 indicates homogenization, while greater than 1 suggests diversification. Through controlled experiments involving 30 GPT-4o agents and nearly 10,000 API calls, the study measured coupling in both text and image communication. Key findings include a significant homogenization effect (CAF=0.803) when agents communicate via text, meaning their outputs become more similar. Conversely, image communication showed a tendency towards diversification (CAF > 1). The research also revealed extreme variation in coupling across different models and that coupling is primarily driven by the immediate prompt context rather than cumulative history, leading to monotonic convergence with continuous consensus. These results firmly establish that LLM agent coupling is a real, measurable phenomenon that can be controlled at the prompt level. This has direct and important implications for how multi-agent LLM systems should be designed to either encourage consensus or maintain diversity in their collective outputs.