Most approaches to agreement have treated the agreement relation as equivalent to a copying or checking relation in which all the agreement features of the agreement source, such as a subject noun, appear on the agreement target, such as a verb. Looking at corpus and typological data, we see that this view of agreement cannot be sustained as a general treatment and, furthermore, it conceals some patterns in the agreement relation. It will be argued that inappropriate data structures have been used in formulating agreement relations and that the long distance consistency intuitively associated with agreement holds at the level of discourse rather than syntax.