One unbelievable action in ten is often enough to break the illusion of autonomous life and make a user uneasy. The more expressive the agent (e.g., Otto graphics vs. AR graphics), and the more sophisticated the information it conveys (e.g., CyberCafe vs. Petz), the more difficult the task of maintaining believability on all fronts. Much of the work discussed in this section is in extracting those components of believable interaction which can transfer from one domain to another, and in creating either theoretical, or pragmatic, tools for simplifying the task for agent-makers. The NYU-IMRPOV and UPENN-JACK groups are building high-level agent interfaces to direct realistically rendered humanoid agents in virtual environments. The Cassell work, now at MIT (but also associated with earlier work at UPENN) is in building avatar agents that have realistic, continuous, autonomous, gesture behavior during discourse, while Mateas is buiding a hybrid semi-autonomous avatar that reports subjective information. The MIT group has studied the role of computer representations of affect, is building emotionally intelligent agents, and has demonstrated agents that operate fully autonomously in navigating virtual environments. Doyle is creating a protocol for extant agents to move from one domain to another, and perhaps from one site to another, while carrying their social intelligence with them, but yet adopting domain intelligence only as needed. The Reeves and Nass group has helped to show that users already treat computer programs as social entities.