This study proposes a dual-agent approach for condition-based maintenance of production lines. The degradation process of production machines is mathematically modeled using an exponential nonlinear Wiener process, and the maintenance decision-making problem is formulated as a Markov decision process (MDP). To enhance decision-making flexibility, a multi-stage maintenance strategy is designed, incorporating perfect maintenance, imperfect maintenance, corrective maintenance, and maintenance timing decision, in addition to conventional do-nothing action. To mitigate the curse of dimensionality, a dual-agent deep reinforcement learning (DRL) framework is proposed, where maintenance decision-making and maintenance scheduling are treated as separate tasks assigned to two agents. Furthermore, a counterfactual reward mechanism is used to optimize reward allocation between the agents. Finally, the experiments on synchronous and asynchronous systems show improved efficiency and reduced maintenance costs compared to traditional methods.