Enhancing the resilience of smart grids is crucial for maintaining the security and reliability of power systems.An adaptive graph attention multi-agent reinforcement learning （AGA-MARL ） method is proposed，through which the learning efficiency and collaborative ability of the system in complex grid environments are improved by an adaptive learning rate and a dynamic task allocation mechanism，thereby enhancing the resilience and interpretability of smart grids.First，adaptive multi-agent deep reinforcement learning （AMA-DRL ） and dynamic spatial-temporal graph convolutional networks （DST-GCN ） are combined to enhance the information interaction among multiple agents and utilize the dynamic graph structure to capture the complex dependencies of the grid system.Second，an interpretability module is integrated to provide more intuitive decision-making explanations by combining attention weights and Shapley value （SHAP）.Finally，the effectiveness of the proposed method is verified through experiments.The research results show that compared with the traditional AMA-DRL method，AGA-MARL performs better in aspects such as grid fault recovery time，system stability，and interpretability.