The article introduces MAVIC, a novel approach to multi-agent reinforcement learning that enhances instruction compliance without sacrificing performance on primary tasks. This method addresses the challenges of inconsistent value estimates when external instructions conflict with ongoing actions in complex environments.