Multi-agent adversarial policy learning is a vital research direction in current artificial intelligence and game theory. However, how to construct efficient and adaptable policy learning methods to cope with dynamic and complex environments is one of the core challenges in multi-agent game research. To address this, this paper proposes an Event Chain-based Multi-Agent Reinforcement Learning method. The core innovation of EC-MARL lies in introducing the event chain mechanism. By capturing the temporal correlations among critical decision-making nodes such as “interception timing, invasion strategy switching, and target route response” during offensive-defensive interactions, it enhances agents’ perception of cooperative/adversarial relationships, thus optimizing policy learning efficiency and balancing offensive-defensive performance. The effectiveness of the proposed method is verified by comparing it with four baseline models. Overall, EC-MARL effectively alleviates the offensive-defensive imbalance, accelerates policy convergence, and demonstrates superior performance in multi-agent adversarial environments.