Global disasters are increasing in scale and frequency, disrupting livelihoods worldwide and intensifying food insecurity. Quantifying how shocks alter agricultural commodity shipping volumes and prices is essential to identify fragilities in trade networks and to design resilient mitigation strategies. This work presents a hybrid multi-agent reinforcement learning (MARL) approach for solving variational inequality (VI) problems that model multi-commodity trade network equilibria. While VIs provide a rigorous framework for capturing interacting supply–demand decisions under network constraints, their solution via MARL is still limited by stability and convergence challenges. Building on an actor–critic backbone, the proposed method integrates a gradient-based learning-rate scheduler, adaptive exploration decay, prioritized replay, and a dual reward that combines individual incentives with centralized feedback to steer agents toward equilibrium. Agents represent supply- and demand-side actors that learn decentralized strategies through repeated interaction in simulated markets. Experiments progressively increase environmental complexity, moving from stable conditions to time-varying prices and disruption scenarios such as route blockages. Results indicate faster, more reliable convergence than baseline MAPPO and MADDPG implementations, alongside robust adaptation under dynamic and adverse conditions. These findings support MARL as a practical tool to simulate economic behavior and optimize decentralized decision-making in complex, networked trade systems.