In this article I describe some of the features of human-human conversation that are being implemented in this new genre of embodied conversational agents. Then I describe an embodied conversational agent that is based on these features. I argue that, because conversation is such a primary skill for humans, and such an early-learned skill (practiced, in fact, between infants and mothers who take turns cooing and burbling at one another), and because the body is so well-equipped to support conversation, embodied conversational agents may turn out to be powerful ways for humans to interact with their computers. However, I claim that in order for embodied conversational agents to live up to their promise, their implementations must be based on actual study of human – human conversation, and their architectures must reflect some of the intrinsic properties found there.