Abstract: The N-1 static security verification is an important part of the power system security and stability analysis. When the system can not meet the requirements of static security, appropriate preventive control measures should be taken. It is the most important preventive control measure to adjust generators' power. The traditional method for the control strategy decision is based on experts' knowledge and experience in combination with digital simulation, which usually takes long time. Deep reinforcement learning, having the advantage of "offline training, online end-to-end strategy formation", is suitable for application in power system preventive control decision. However, how to reduce the search space and improve the training speed is a problem that needs to be solved. In this paper, a double-agents generator power adjustment method is proposed based on soft actor-critic (SAC) deep reinforcement learning algorithm. Considering that power transmission network has the feature of PQ decoupling, a cooperative double-agents structure with centralized training is designed. In the structure, two agents take on the tasks of active power adjustment and voltage regulation of generators respectively and cooperate with each other according to the N-1 verification of the whole power system under different operation states, and the search space for the control strategy can be reduced and the model stability can be improved effectively. Case studies on the IEEE 39-bus system show that the proposed method can quickly and effectively generate preventive control strategies under various operation states, which verifies the effectiveness of the proposed method.