Task scheduling is key to performance optimization and resource management in cloud computing systems. Because of its complexity, it has been defined as an NP problem. We introduce an online scheme to solve the problem of task scheduling under a dynamic load in the cloud environment. After analyzing the process, we propose a server level agreement constraint adaptive online task scheduling algorithm based on double deep Q-learning (SLA-DQTS) to reduce the makespan, cost, and average overdue time under the constraints of virtual machine (VM) resources and deadlines. In the algorithm, we prevent the change of the model input dimension with the number of VMs by taking the Gaussian distribution of related parameters as a part of the state space. Through the design of the reward function, the model can be optimized for different goals and task loads. We evaluate the performance of the algorithm by comparing it with three heuristic algorithms (Min-Min, random, and round robin) under different loads. The results show that the algorithm in this paper can achieve similar or better results than the comparison algorithms at a lower cost.