A cluster of computers is required to execute large-scale multi-agent. However, such execution incurs an inter-node communication overhead because agents intensively communicate with other agents to achieve common goals. Although a number of dynamic load balancing mechanisms have been developed, these mechanisms are not scalable in multi-agent applications because of the overhead involved in analyzing the communication patterns of agents. This paper proposes two scalable dynamic agent distribution mechanisms; one mechanism aims at minimizing agent communication cost, and the other mechanism attempts to move agents from overloaded agent platforms to lightly loaded platforms. Our mechanisms are fully distributed algorithms and analyze only coarse-grain communication dependencies of agents, thus providing scalability. We describe the results of applying these mechanisms to large-scale micro UAV (Unmanned Aerial Vehicle) simulations involving up to 10,000 agents.