Despite many years of work on multi-agent platforms development, the problem of their performance remains unsolved. Even most mature solutions still suffer from significant limitations in terms of message-passing services and the number of agents which can work simultaneously. Large overheads are caused mostly by unsuitability of technology used for implementing the agent platform itself. In this paper several measures of an agent platform performance are proposed. Two popular agent platforms are thoroughly tested and compared to Erlang technology. The results of the experiments show that different approach to the problem of creating agent platform can give significantly better performance.