This thesis presents the development and evaluation of a Double Deep Q-learning Network (DDQN) agent for addressing the Dynamic Job-Shop Scheduling Problem (DJSP) in high-utilization systems. The DDQN agent is designed to optimize job scheduling dynamically, taking into account stochastic arrivals and stochastic jobs attributes. By employing separate networks for action selection and evaluation, the DDQN mitigates overestimation bias, enhancing the stability and accuracy of the learning process. Preliminary results indicate that the DDQN agent performs well in highly utilized systems, demonstrating significant promise in optimizing scheduling efficiency, though its performance is comparable to some well-tailored traditional heuristics. However, its performance in less utilized systems remains less effective, suggesting room for further refinement.