Cyber-Physical Systems (CPSs), such as self-driving vehicles, have become increasingly complex, involving the use of multiple control tasks to meet the desired safety requirements. One challenge is to lower costs by reducing hardware requirements for executing these tasks. Traditionally, meeting task deadlines was considered to be a measure of safety, but in recent times, there has been work in synthesizing schedules where tasks can miss deadlines such that a system-level notion of safety is still satisfied. However, such approaches are inflexible, as the task schedules may need to be re-synthesized when new tasks are introduced. We propose a model-less Reinforcement Learning (RL) motivated approach for schedule synthesis, where tasks can dynamically be added or removed. This RL strategy is guided by a quantifiable system-level safety property. It offers a more flexible approach to improving resource efficiency, while maintaining safety. We illustrate its benefits using both deterministic and stochastic versions of our proposed schedules.