In this paper, a novel bandwidth negotiation mechanism is proposed for massive devices wireless spectrum sharing, in which individual device locally negotiates bandwidth usage with neighbor devices and globally optimal spectrum utilization is achieved through distributed decision-making. Since only sparse feedback is needed, the proposed mechanism can greatly reduce the signaling overhead. In order to solve the distributed optimization problem when massive devices coexist, mean field multi-agent reinforcement learning (MF-MARL) based bandwidth decision algorithm is proposed, which allow device make globally optimal decision leveraging only neighborhood observation. In simulation, distributed bandwidth negotiation between 1000 devices is demonstrated and the spectrum utilization rate is above 95%. The proposed method is beneficial to reduce spectrum conflicts, increase spectrum utilization for massive devices spectrum sharing.