In this paper, we examine the effectiveness of the social distance between a virtual agent and users, and the gaze instruction using a display that can be viewed stereoscopically without using any wearable devices. An actual robot cannot always maintain an appropriate interpersonal distance, through nonverbal gestures owing to its limited range of motion. Because large movements may harm humans, the nonverbal gestures of robots are limited in the real world. In this work, 14 participants were asked how far they wanted to move from a robot posing as a museum guide agent. They were also asked to identify the point at which they felt the agent was gazing. There was a significant distance between the initial position and the position to which the participants moved in the first task under two-dimensional (2D) and three-dimensional (3D) scenarios. The participants moved a significant distance in the first task. In the gaze estimation task, however, the error between the 3D and 2D evaluations was significantly lesser, at a point far from the agent.