Group intelligence is becoming increasingly important in modern military systems, as it can significantly reduce operational risks and improve overall mission effectiveness. Coordination and cooperation among agents are critical for enabling a group to achieve comprehensive advantages in complex battlefield environments. In this paper, a cooperative decision-making model for multi-agent reinforcement learning, termed TransGMix, is proposed. The model consists of two main components: a decision model and a mixing model. In the decision model, a Transformer is introduced to encode the agents’ local observations, through which individual policies are generated. Furthermore, a decision experience update mechanism is designed to enhance temporal consistency by incorporating historical decision information, thereby improving the stability and accuracy of the current decision process. In the mixing model, a hypernetwork is constructed by integrating graph neural networks and multilayer perceptrons. Through value decomposition, the proposed structure facilitates information integration among agents and promotes cooperative optimization at the group level, enabling a unified transition from individual rationality to collective coordination. Experiments conducted in multi-scenario simulation environments, such as the StarCraft Multi-Agent Challenge, demonstrate that TransGMix achieves an average win rate of 94.5%, indicating that the proposed method can effectively capture emergent cooperation and global optimization capabilities in multi-agent systems.