The goal of this project is to create an agent using traditional reinforcement learning methods that can play a number of early combats in the card game Slay the Spire to the same level of skill as a human player with a few dozen hours of experience. In particular this project utilizes a direct policy search model to simplify the state-action space such that the agent isn’t overwhelmed with data and can be trained significantly faster. In the case of early fights tested, the agent made near optimal or better moves in the vast majority of cases encountered, indicating that this method could be successfully extended to navigate more complex situations.