In the era of the Internet of Everything, the traditional computing par­adigm gradually fails to satisfy the tasks with high quality of service require­ments, and in-network computing gradually enters the field of vision, providing a solution idea for the case of insufficient computing network resources. In-net­work computing involves scheduling computational tasks on the host computer to be executed on network nodes, utilizing network devices such as switches. It simultaneously performs online processing of data during transmission to reduce communication delay, lower energy consumption, and improve overall compu­tational efficiency. This paper studies the joint planning method for the deploy­ment and routing of in-network computing tasks, incorporating knowledge from both in-network computing and reinforcement learning. We also examine exist­ing task deployment and routing methods. We design a joint planning model for the deployment and routing of in-network computing tasks and propose a joint planning algorithm (MDL-TR) based on a message passing neural network and deep reinforcement learning to achieve optimal task scheduling. Finally, the ef­fectiveness of the proposed model is verified through simulation experiments.