A performance model for the interaction of agents in mobile agent systems is presented. Two interaction models, namely the remote procedure call and the agent migration are considered. Performance models for a single interaction are introduced, which are then used to derive a performance model for a sequence of interactions. This performance model can be used to evaluate the performance of any possible behaviour of an agent in a given scenario. The application of the performance model for a typical scenario in mobile computing shows that the optimal behaviour of an agent is achieved by a mixed sequence of remote procedure calls and agent migrations. The performance model is validated by measurements of interactions of real agents in the mobile agent system Mole.