Multi-agent trajectory prediction is a critical task in autonomous driving and robotic control, requiring precise modeling of complex interactions among agents and their dynamic relationships with the environment. In this paper, we introduce TrajFusionNet, a novel framework that integrates spatio-temporal feature fusion with consistency constraints to improve prediction accuracy. Specifically, our method employs a multi-scale temporal convolutional network (M-TCN) to capture long-term dependencies in agent motion, while a spatio-temporal graph network models structured interactions among agents. To enhance the prediction process, we propose a gating mechanism that dynamically fuses spatial and temporal features, ensuring a more effective representation of motion patterns. Furthermore, we design a self-consistency constraint module to enforce prediction stability under spatio-temporal perturbations, reducing the impact of noise and minimizing prediction inaccuracies. Extensive experiments on the ETH/UCY and Stanford Drone Dataset (SDD) demonstrate that TrajFusionNet consistently outperforms state-of-the-art methods, achieving superior accuracy in complex scenarios. This research offers a high-precision, high-stability trajectory prediction solution with broad applications in autonomous driving, multi-robot coordination, and intelligent transportation systems.