In multi-agent systems, roles provide an apriori decomposition of the task based on the interactions of the agents. Roles are instantiated by the agents’ actions and/or messages. Role allocation has emerged as a key computational tool in the coordination of intelligent agents who are flexible enough to take on any role. It is in this context that some rethinking about roles and their representation is needed. Illustrations in the game of Go, a game of strategy, highlight the key issues.