Abstract: Aiming at the problems of no measured operation data, difficulty in verifying collaborative scheduling schemes, and inability to quantify safety risks in the feasibility study stage of open-pit mine clusters, a technical framework combining graph networks, graph convolutional networks, and multi-agent simulation is proposed for collaborative intelligent scheduling and safety decision-making. The mining, transportation, dumping, and monitoring units of the mine cluster are abstracted as graph nodes with topological relationships. A single-layer graph convolutional network (SGCN) is employed to extract features and perform spatial encoding of the system's global state as part of the state representation. A multi-objective reward mechanism is constructed to balance production efficiency, collaborative safety, and resource utilization. Driven by parameterized rules, the agents in the simulation make action decisions based on predefined engineering rules rather than relying on features extracted by the graph convolutional network, ultimately enabling collaborative strategy deduction and risk pre-assessment under conditions of no mine operation data. A case study is conducted on the collaborative mining of two adjacent open-pit mines. The results show that in this collaborative mining scenario, the task completion rate reaches 72.9%, the internal dumping ratio is 79.1%, and safety violations remain at a controllable level. The simulation results preliminarily validate the logical feasibility of the model under specific assumptions, providing a theoretical reference for scheme comparison and safety pre-judgment in the early-stage joint mining of mine clusters. Finally, comparative validation, parameter sensitivity analysis, and a summary of the method's advantages, limitations, and future research directions are presented.