This paper proposes a GNN-enhanced multi-agent deep reinforcement learning framework for antijamming routing in mobile ad hoc networks. The approach combines temporal graph representations with centralized training and decentralized execution, together with a learning-based jammer to model adversarial interference. Simulations (10 runs, mean ± standard deviation) show that the proposed method converges in approximately 100 episodes and achieves a throughput of 2.6±0.05 Mbps, improving over QMIX (2.3 Mbps) and DQN (2.0 Mbps). It attains a packet delivery ratio of 0.85±0.02, reduces end-to-end delay to 65±5 ms, and lowers energy consumption to 70±5 J under strong jamming. The performance gains are consistent across mobility levels and interference intensities, with reduced variance compared to baseline methods. Ablation results indicate that temporal graph modeling provides the largest contribution, while adversarial training improves robustness. These results demonstrate that the proposed framework enables stable and efficient routing in dynamic and non-stationary wireless environments, with practical scalability due to decentralized execution.