If the audience stages the scene, who directs the story?

Storycaster asks a different question of Agentic Media: what if the interface is the room?

Presented at CHI 2026, Storycaster is a generative-AI CAVE system that turns a physical room into a responsive storytelling environment. Live camera feeds drive cylindrical projections across the walls, while the setup keeps participants aware of the real space instead of enclosing them in a headset. 1

Its narrator agent guides voice-driven co-creation. Each scene can combine generated imagery, ambient audio, dialogue, and edits that turn a physical object into a virtual story element. The participant supplies the direction; the agent coordinates the scene. 2

The handoff is the interesting part. Storycaster organizes the experience as setup, confrontation, and resolution. At each transition, participants can continue, change an object, choose a suggested direction, or let the Narrator decide. Agency is therefore a turn-taking rule, not a checkbox in a settings panel. 3

The prototype also shows where the medium strains. In a study with 13 participants, the narrator and audio were rated the most impactful elements, while latency and image resolution were flagged for improvement. The paper reports roughly 7.3 seconds for image generation and 12 seconds for ambient-audio generation, alongside uneven object detection and the limits of a fixed calibrated room. 2

Discussion: If an agent controls pacing, scene generation, and what counts as an identifiable object, is the participant a co-author, or a guest inside the system's rules? What should be inspectable: the prompt, the scene state, the generated assets, or all three?