When coordinating decisions in latency-sensitive 6G networks, distributed agents must act without exchanging messages at runtime. A well-known obstacle arises from the CHSH (Clauser–Horne–Shimony–Holt) inequality, which caps classical coordination accuracy at \(75\%\) for the standard non-local game under a strict no-signaling protocol. We introduce eMARL (entanglement-assisted Multi-Agent Reinforcement Learning), a learning framework in which agents share pre-distributed Bell pairs and learn measurement policies to surpass this classical ceiling. Across \(n=50\) independent training runs, eMARL reaches a win rate of \(85.29\%\) and a CHSH parameter \(S = 2.817\), while transmitting zero bits during execution. Meanwhile, bandwidth-limited learned-communication baselines such as CommNet (\(71.6\%\)) and TarMAC (\(71.7\%\)) suffer from training instability and fall short of even the classical optimum. We characterize performance under both depolarizing and photon-loss channels. The depolarizing analysis fixes a fidelity threshold for automatically falling back to classical policies; the loss analysis distinguishes the two operating regimes of photonic networks (heralded post-selection and unheralded amplitude damping), the former preserving the conditional quantum advantage but paying a \((1-\eta )^2\) throughput penalty per round. We close the loop with a simplified Cell-Free MIMO (Multiple-Input Multiple-Output)-inspired beamforming task, where eMARL recovers \(95.3\%\) of centralized throughput with zero runtime signaling. Taken together, these findings position entanglement as a promising coordination resource for CHSH-type distributed problems in principle.
