Matchmaking will be an important component of future agent and agent-like systems, such as the semantic web. Most research on matchmaking has been directed toward sophisticated matching of client requirements with provider capabilities based on capability descriptions. This is a vital mechanism for conducting matchmaking, but ignores the likelihood that in practise, and for various reasons, capability descriptions will not fully characterise the interaction behaviour of agents. This problem is further compounded in systems with many interacting agents, all of which have idiosyncrasies. As in everyday life, some groupings of agents will be more effective than others, regardless of their individual competencies or suitability to the task. The quality of the interaction between agents is a crucial factor. Using the incidence calculus and the lightweight coordination calculus, we show that we can easily implement matchmaking agents that will learn from experience how to select those groups known to inter-operate well for particular tasks.