In this proposal, I outline a generator conceived of as part of a general intelligent agent. The generator's task is to provide the overall system with the ability to use communication in language to serve its purposes, rather than to simply encode information in language. This requires that generation be viewed as a kind of goal-directed action that is planned and executed in a dynamically changing environment. In addition, the generator must not be dependent on domain or problem-specific information but rather on a general knowledge base .that it shares with the overall system. These requirements have specific consequences for the design of the generator and the representation it uses. In particular, the text planner and the low-level linguistic component must be able to interact and negotiate over decisions that involve both high-level and low-level constraints. Also, the knowledge representation must allow for the varying perspective that an intelligent agent will have on the things it talks about; the generator must be able to appropriately vary how it describes things as the system's perspective on them changes. The generator described here will demonstrate how these ideas work in practice and develop them further.