The majority of Internet applications requires interaction between various entities through the network, in order to exchange data and to distribute tasks. The client/server model where exchanges are given by distant interactions is the most used model. It has the disadvantage of increasing network traffic by exchanging intermediary information. In this paper we show that in some cases it is better to send the code to the server and to work locally. We study two models of communication (the client/server and the mobile agent) and propose a hybrid one, where agent uses the two models and chooses the best one according to the characteristic of the network. Interaction between different application entities is formulated by a probabilistic model. The Markov decision processes (MDPs) are used to calculate the optimal policy for agent displacement. This policy is applied by the agent in order to decrease network traffic.