The missing of an appropriate semantics of agent communication languages is one of the most challenging issues of contemporary AI. Although several approaches to this problem exist, none of them is really suitable for dealing with agent autonomy, which is a decisive property of artificial agents. This paper introduces an observation-based approach to the semantics of agent communication, which combines benefits of the two most influential traditional approaches to agent communication semantics, namely the mentalistic (agent-centric) and the objectivist (i.e., commitment- or protocol-oriented) approach. Our approach makes use of the fact that the most general meaning of agent utterances lays in their expectable consequences in terms of agent actions, and that communications result from hidden but nevertheless rational and to some extent reliable agent intentions. In this work, we present a formal framework which enables the empirical derivation of communication meanings from the observation of rational agent utterances, and introduce thereby a probabilistic and utility-oriented perspective of social commitments.