Human languages use a wide range of grammatical categories to constrain which words or phrases can fill certain slots in grammatical patterns and to express additional meanings, such as tense or aspect, through morpho-syntactic means. These grammatical categories, which are most often language-specific and changing over time, are difficult to define and learn. This paper raises the question how these categories can be acquired and where they have come from. We explore a usage-based approach. This means that categories and grammatical constructions are selected and aligned by their success in language interactions. We report on a multi-agent experiment in which agents are endowed with mechanisms for understanding and producing utterances as well as mechanisms for expanding their inventories using a meta-level learning process based on pro- and anti-unification. We show that a categorial type network which has scores based on the success in a language interaction leads to the spontaneous formation of grammatical categories in tandem with the formation of grammatical patterns.