Dense industrial IoT (IIoT) deployments operate in congested 2.4 GHz environments where heterogeneous wireless technologies, such as IEEE 802.15.4, BLE, and Wi-Fi, compete for spectrum. This coexistence produces volatile interference patterns characterized by RSSI fluctuations, bursty collisions, and rapidly changing traffic loads. Traditional MAC-layer protocols are unable to adapt to such dynamics, resulting in degraded reliability and excessive latency. While recent works explore AI for wireless context, they often lack integration between inference and control or fail to address resource constraints. This paper proposes "Learning to Listen"(L2L), a two-stage architecture combining a Graph Neural Network (GNN) to infer real-time interference patterns using only local features, and a decentralized Multi-Agent Proximal Policy Optimization (MAPPO) scheduler that optimizes channel, power, and timing. We demonstrate a TinyML deployment on ESP32-C3 nodes with fewer than 100k parameters. Evaluating the system via ns-3 simulations (800 nodes) and a 20-node testbed, we achieve PDR ¿ 99%, p95 latency ¿ 100 ms, and 20-30% energy reduction compared to CSMA/CA. Furthermore, we integrate SHAP and EBM to provide operator-interpretable explanations for scheduling decisions.
