Talking is a complex task in which a speaker must transform a communicative intent into a series of motor actions in real time. Empirical research on language production has revealed that it is a largely incremental process: speakers do not start with a full plan for an utterance, but instead plan it as they go. I present a theory of incremental language production based on a combination of information theory and control theory, where the choice to produce a particular word in context is determined by a policy that maximizes communicative reward subject to a channel capacity constraint on cognitive control. I show that the theory captures human data on errors in word choice, word order preferences based on the “accessibility” of words, and disfluencies.