You got it backwards. RL is long running inference plus training at the end. They let the Agent spin for quite a while, then they calculate the reward afterwards and modify the parameters accordingly.