With the rapid advancement of artificial intelligence, multi agent reinforcement learning (MARL) has been widely adopted for tasks such as unmanned vehicle formation, robot cooperation, resource scheduling, and collaborative decision-making due to its strong capability to model real world scenarios. However, packet loss in non-stationary communication environments remains a primary obstacle to effective multi agent coordination. To address this issue, this paper introduces a communication mechanism within the QMIX framework and proposes a packet loss information reconstruction technique. By learning the loss function between real messages and predicted messages, it is used to predict messages lost due to unstable communication environments, thereby enhancing the robustness of the model. Experimental results on the SMAC benchmark demonstrate that our approach achieves an average improvement of approximately 5%~15% in collaborative decision-making performance over baseline algorithms and exhibits marked gains in communication resilience. Provide practical and feasible solutions for future research methods related to multi agent systems.