Offshore natural gas compression systems are characterized by strong hydraulic coupling, nonlinear behaviour, and strict safety constraints, particularly in high-CO2 production environments. Conventional decentralized PID control with anti-surge protection ensures reliable local regulation but often leads to poor plant-wide coordination and persistent offsets when multiple compression trains, recycle loops, and separation units interact dynamically. Although multivariable control strategies such as model predictive control can address these issues, their industrial application remains limited by modeling effort, computational demand, and robustness concerns. This work presents a hybrid supervisory control framework in which reinforcement learning (RL) augments an existing PI-based architecture for an offshore gas compression system with membrane-based CO2 separation. A Proximal Policy Optimization (PPO) agent is trained on a dynamic digital-twin model of export, CO2, injection, and bypass compression trains with shared headers, recycle flows, and thermal constraints. The RL agent acts exclusively at the supervisory level, providing bounded incremental adjustments to pressure and bypass setpoints, while all regulatory PI loops, anti-surge protections, and safety logic remain unchanged. Closed-loop simulations under representative disturbance scenarios show that the RL-supervised architecture significantly improves plant-wide coordination. Compared to fixed-setpoint PI operation, total tracking error (IAE) and time-weighted error (ITAE) are reduced by approximately 54% and 56%, respectively, using smooth, low-magnitude supervisory actions. In addition, RL achieves performance close to that of a nonlinear MPC benchmark based on a random-shooting, MPPI (Model-Predictive Path Integral) formulation, while requiring substantially lower online computational effort. The results demonstrate that RL can be effectively integrated as a supervisory layer in offshore gas compression systems without replacing established industrial control infrastructure, although additional safety mechanisms are still required to ensure stricter constraint satisfaction.