With the rapid advancement of information and communication technology, there arises an increasing demand for heightened computing capabilities to facilitate a myriad of intelligent services. The computing power network (CPN) amalgamates an array of resources including cloud and edge devices to enable collaborative sensing, routing, and scheduling. This integration significantly amplifies the utilization of computing network resources while concurrently augmenting service quality. In the realm of multi-node distributed joint computing within CPNs conventional scheduling methodologies frequently neglect the interplay between computing and the network, resulting in suboptimal task scheduling. To address this issue, this study proposes a delay-optimized task scheduling mechanism based on deep reinforcement learning. Simulation results demonstrate that this framework achieves faster convergence speed, particularly under dual-depth networks, outperforming other algorithms in reducing system delays, energy and enhancing scheduling efficiency.