Information Homogenization Induces Herding in Retrieval-Augmented LLM Agent Markets

Large language model (LLM) agents augmented with retrieval mechanisms are increasingly deployed in collective decision-making settings such as trading, coordination, and automated analysis. Prior work on retrieval-augmented generation (RAG) has primarily focused on improving factual accuracy and robustness at the individual agent level, while implicitly assuming that shared retrieval is benign or beneficial. However, classical studies in economics and complex systems suggest that information homogenization can induce herding and systemic risk, a perspective that has been largely absent from current LLM agent research. In this work, we study how information overlap in RAG affects collective behavior among LLM agents. We define a controlled overlap ratio $\rho$ that governs the fraction of shared retrieval documents across agents, and construct a multi-agent trading framework in which agents operate on identical market data under fixed, conservative risk controls and without communication. This design isolates information structure as the sole coupling mechanism. Across minute-level cryptocurrency backtests, we observe a monotonic increase in behavioral homogeneity as $\rho$ increases from 0 to 1, with mean pairwise action correlation rising from 0.04 to 0.41 and herding ratios increasing accordingly. Higher overlap also leads to increased system-level position volatility and stronger action–market coupling, while average individual profitability does not systematically improve. These results demonstrate that information homogenization alone can induce herding and systemic fragility in retrieval-augmented LLM agent systems, highlighting information diversity as a critical design consideration.