Efficient congestion management in Software-Defined Networks (SDNs) remains a significant challenge due to dynamic traffic patterns and complex topologies. Conventional congestion control techniques based on static or heuristic rules often fail to adapt effectively to real-time network variations. This paper proposes a data-driven framework based on Multi-Agent Reinforcement Learning (MARL) to enable intelligent, adaptive congestion control in SDNs. The framework integrates two collaborative agents: a Congestion Classification Agent that identifies congestion levels using metrics such as delay and packet loss, and a Decision-Making Agent based on Deep Q-Learning (DQN or its variants), which selects the optimal actions for routing and bandwidth management. The agents are trained offline using both synthetic and real network traces (e.g., the MAWI dataset), and deployed in a simulated SDN testbed using Mininet and the Ryu controller. Extensive experiments demonstrate the superiority of the proposed system across key performance metrics. Compared to baseline controllers, including standalone DQN and static heuristics, the MARL system achieves up to 3.0% higher throughput, maintains end-to-end delay below 10 ms, and reduces packet loss by over 10% in real traffic scenarios. Furthermore, the architecture exhibits stable cumulative reward progression and balanced action selection, reflecting effective learning and policy convergence. These results validate the benefit of agent specialization and modular learning in scalable and intelligent SDN traffic engineering.