As new transport protocols are being proposed and standardized, the choice of the best communication service to be used by applications for delivering their data when distributed is becoming too complex. Application developers need much knowledge on "how the protocol works" to decide whether or not it can be used to fulfill their requirements. Moreover, the performance of the service provided by a given communication protocol is highly dependent on the network context. The Autonomic Transport Protocol presented in this paper is aware of the application requirements and uses learning techniques to adapt the service it provides to best satisfy these requirements as the network conditions vary.