This paper examines theoretical analyses of vowel harmony in the light of data from Warlpiri, a language of Central Australia. Modern analyses account for vowel harmony in terms of a constraint requiring feature agreement within a particular domain (e.g. Backovic, 2000; Lombardi, 1996). These analyses predict that exclusively suffixing languages should exhibit only rightwards harmonies. They also predict that all disharmonic sequences within a particular domain should be equally disfavoured. Warlpiri is of interest for two reasons. Firstly it is an exclusively suffixing language, but it shows both leftwards and rightwards harmonies. Secondly, it permits one disharmonic sequence within the lexicon, but not its converse. We show that the leftwards harmony is not an exception to the predictions of modern analyses, as it is morphologically and not phonologically motivated. However, we show that agreement constraints cannot account for the evidently related harmony constraints on the structure of the lexicon in Warlpiri. Rather, these must be accounted for by constraints against specific disharmonic sequences. Under this ‘anti-disagreement’ analysis the structure of the lexicon follows from the grammar.