Abstract. The paper presents the multi-agent, parallel computing system (MAS) composed of a platform of software servers and a set of computing agents. The generic actions of agents and the system government are so designed that it can perform irregular concurrent genetic computations in heterogeneous computer network with a number of computation nodes and connection topology varying in time. The effectiveness of MAS solution is discussed in terms of average migration and communication overheads. Additionally, the MAS system with autonomous, diffusionbased scheduling is compared with low-level distributed implementation, which utilizes the centralized greedy scheduling algorithm.