Cognitive Heterogeneity and Behavioral Biases in Multi-Stage Supply Chains: Evidence from LLM-Based Agent Simulations

Modeling the cooperation and coordination of generative agents in complex, multi-round dynamic decision-making scenarios remains a central challenge for artificial intelligence and operations management. While traditional behavioral experiments in multi-stage supply chains have revealed how cognitive biases lead to systemic inefficiencies, such studies often face high costs, limited scalability, and difficulties controlling for individual heterogeneity. Recent advances in Large Language Models (LLMs) offer a scalable and reproducible alternative through agent-based simulations. In this paper, we introduce a novel experimental paradigm using agents powered by the DeepSeek and ChatGPT model series to simulate multi-stage supply chain management. Our study is grounded in a Hierarchical Reasoning framework, which informs our analysis of cognitive heterogeneity among agents. Through a series of replicated and statistically-validated simulations, we systematically vary agent sophistication across supply chain tiers to investigate the impact of this cognitive diversity on collective outcomes. The results show that LLM agents can display myopic and self-interested behaviors, which intensify supply chain inefficiencies. We find, however, that information sharing serves as an effective intervention to mitigate these adverse effects. This work extends traditional behavioral research methods and provides new insights into the emergent dynamics of AI-enabled organizations. Our findings highlight both the promise and the limitations of using LLM-based agents to study complex human-like decision-making in supply chain contexts.