This paper investigates the problem of multi-user computation task offloading and resource allocation for smart power systems. Considering the heterogeneous nature of devices and the latency-sensitive, computation-intensive tasks of critical machine-type communication devices, we formulate a joint optimization problem to minimize their overall computational load under quality of service constraints. To address the challenge, a distributed algorithm based on multi-agent deep Q-networks is developed, enabling agents to learn dynamic offloading and resource strategies through environmental interactions. Simulation results show that the proposed method significantly reduces system load compared to existing algorithms. This work demonstrates the potential of M2M communication for energy-efficient task offloading and provides a viable low-latency solution for collaborative computing in power IoT networks.