A smart Load Balancing (LB) policy based on Graph Convolutional Multi-Agent Reinforcement Learning (GC-MARL) is proposed to improve load balancing in networks beyond what can be realized by traditional methods and state of the art machine learning based approaches. GC-MARL models the network as a graph and derives through a graph convolutional method the policy that splits traffic flows across end-to-end candidate paths while meeting application QoE requirements. The proposed method uses the throughput and the delay, observed at the network level, as the key performance indicators embedded in the reward expression as opposed to observing QoE at the application level. The results confirm the effectiveness of the proposed solution in terms of KPIs (such as throughput, delay, jitter, packet loss), and KQIs (such as QoE, average video bitrate, stalling, etc…).