In the human-computer interaction (HCI) practice, there have been increasing interests in systems consisting of multiple conversational agents (CAs). However, existing studies pointed out that the interactions with multiple CAs require a high level of users’ mental efforts. To cope with this issue, we introduce a mediator agent (MA): a single CA which mediates interactions between a user and multiple CAs. Although there are existing studies that imply the potential of a CA as a mediator in a multi-party context, understanding of user expectations on MA is still lacking. In this work, we conducted an improvisational acting-based Wizard of Oz study to explore user experience (UX) of MA-mediated interactions with multiple CAs. Twelve participants experienced interactions mediated by a virtual MA and interviewed about their expectations on a MA. Qualitative analysis of interview data revealed user expectations on a MA, including user-preferred types of agent-mediation such as confirmation-based mediation and transparency-oriented mediation.