Abstract: Optimizing the opening settings of an intelligent sliding sleeve completion system in horizontal wells is a challenging problem due to its multivariable, strongly coupled, and nonlinear nature. In this study, a coupled flow model for the sliding sleeve-reservoir system in a horizontal well is developed and validated against numerical simulation, with relative errors within 12%. Based on this model, a reinforcement learning framework guided by Markov decision theory is adopted. By incorporating prioritized experience replay (PER) into the multi-agent deep deterministic policy gradient (MADDPG) framework, a PER-MADDPG algorithm is proposed to optimize multistage sliding sleeve completion in horizontal wells. The results indicate that, compared with DDPG and conventional MADDPG, PER-MADDPG achieves higher rewards with lower training loss and improved convergence. The optimized strategy increases cumulative oil production by 10.7% and reduces the water cut by 6.4% after two years of production. By selectively prioritizing and replaying informative experiences, PER-MADDPG significantly enhances learning efficiency and sample utilization, providing theoretical support for real-time multi-agent control in horizontal well production.