The proposed approach to avoiding catastrophic failures in interconnected power systems uses a defeusive system based on multiagent system technologies. Because of their dynamic environments, these agents need to automatically adjust their decision criteria. The method of reinforcement learning with temporal differences can provide a suitable basis for the adaptation. An appropriate convergence criterion is derived and an application of the method to load shedding is demonstrated. A dynamic simulation involving a 179-bus power system is used to validate the method.