In bulk power systems， cascading failures triggered by normal events may lead to large blackouts. For online security control in large scale power grids， a real-time control strategy based on multi-agent system to prevent cascading trips is presented， and the principle of load shedding and generator tripping is also given. An agent is placed at each bus of the power grid differing from traditional distributed method that takes subarea or substratum as an agent. Each agent employs model predictive control to solve its local problem based on locally available data and the global solution can be acquired with agents' cooperation. Furthermore， a filtering method for control variable and constraint set is employed， and linear programming method is used in every varying receding-horizon. The data for next varying receding-horizon control can be acquired from the measured data of the real power grid. Simulation results of IEEE 118-bus system verify the fleetness and efficiency of this method.