Today’s agile production systems face an ever-increasing complexity due to individualized mass production, a volatile customer demand and dynamic events which disrupt the baseline schedule. Robustness and optimality of the schedule affect the efficiency of the production systems, which in turn negatively affects resource consumption and waste generation. To achieve near-optimal performance, while adhering to due dates and robustly handling disturbances, predictive-reactive scheduling is often proposed. In the current state-of-the-art research on predictive-reactive scheduling there still exist some open issues. One issue to address is a current deficit in generalization, that is, when handling new, previously unseen data. Arguably the main culprit is the reactive (re)scheduling component. Designing such highly complex systems imposes many difficulties. We identify existing problems which ultimately limit the applicability of predictive-reactive scheduling and investigate and evaluate a design approach for predictive-reactive scheduling. We propose predictive-reactive scheduling, which uses match-up scheduling and promising decentralized multi-agent reinforcement learning, which allows for fast and flexible real-time decision-making. The aim of this study is to provide insights and improve the applicability of research on predictive-reactive scheduling.