This thesis focuses on multi-agent reinforcement learning (MARL) in dynamic and partially observable environments, where decentralized coordination is essential. Traditional CTDE (Centralized Training with Decentralized Execution) approaches assume perfect communication during training, an unrealistic assumption in real systems subject to communication and resource constraints. The objective is to develop a decentralized, hierarchical, and transferable learning framework that allows agents to: communicate adaptively, coordinate their behaviors despite limited observability, and transfer their skills to new tasks, agent configurations, or communication structures, with a view to continuous learning.