This work introduces a novel decision-making framework for autonomous ego vehicles operating at unsignalized intersections under random task scenarios. The proposed framework exploits the power of deep reinforcement learning with the SAC algorithm to address challenging issues, including environmental uncertainty and multitasking complexity. It uses a novel mix-attention network for selective processing of important environmental data, and the formulation of unique state input captures driving tasks. Besides, an improved experience replay mechanism accelerates and stabilizes training. All of these features ensure improved safety and navigation efficiency. Some advantages of the proposed system include superior collision avoidance, adaptability to diverse scenarios, and optimized decision-making under dynamic conditions. Simulation results have validated its effectiveness, where it has shown significant improvements over baseline methods. This study addresses some of the most critical challenges associated with the navigation of autonomous vehicles but also provides a basis for integrating connected vehicle technologies into smarter and safer transportation systems.