Prerequisite relations identify the underlying dependencies among concepts, which are crucial for many intelligent applications in education. Existing neural network-based methods for prerequisite prediction heavily based on labeled data and lack cross-domain adaptability. In this paper, we explore the potential of Large Language Models for prerequisite prediction. More specifically, we propose AutoPRE, a novel LLM-based multi-agent framework to overcome these limitations. AutoPRE introduces three roles: Explainer, Determiner, and Criticizer, around which we design three interaction strategies: Decomposition, Voting, and Confrontation, including 11 methods to improve prediction accuracy. Evaluated on two public datasets, AutoPRE surpasses state-of-the-art neural and LLM-based baselines, achieving a 13.04% improvement in prediction accuracy. AutoPRE’s ability to adapt across domains without extensive labeled data provides a new approach to knowledge modeling in education. This work showcases the potential of LLM-based multi-agent systems with hierarchical interaction designs, paving the way for advanced intelligent tutoring systems.