To effectively support interaction with humans, embodied agents should be implemented using cognitively plausible models such as task models. Even so, these embodied systems should not bear the entire burden for making domain-specific interactions cognitively tractable for humans. Further, embodied systems need not model their users directly; only the necessary and relevant consequences of user state changes or actions should be modeled. The remaining gaps between an embodied system’s implementation and the need to support human interaction should be addressed by human-centered software. We describe the use of liaison agents to bridge these gaps.