Predicting AI Agent Behavior through Approximation of the Perron-Frobenius Operator