OpenAI Experiments - Room Resonance

Daan worked a short FIELD.io sprint (late April to May 2026) for OpenAI's Experiments team, their internal prototyping group, buildingRoom Resonance(also called Spatial Rig): an ambient agent that senses the atmosphere of a physical room and responds to it, rather than reporting on it. The direction was inspired by the Pensieve from Harry Potter, an object that reads the overall vibe of a space, and it was one of several concepts explored in the sprint, but the one Daan built out furthest as working infrastructure.

Room Resonance senses a room's conversational atmosphere through sound energy, temperature, presence, and simple environmental change, and interprets that signal into a small set of bounded group-level states (unknown,stalled,building,coherent,fragmented,intense). Rather than surfacing those states as a readout, the agent acts on the room itself, shifting ambient media, generative sound, or light to match. The one boundary that shaped every decision: this could never become a team score, a diagnostic dashboard, or a productivity readout of the people in the room, only ambient, reversible, peripheral response.

Daan prototyped the sensing and actuation layer across several parallel tracks: a Chromecast (Cast) receiver wired to a NATS message bus for ambient media control, an evolving generative soundscape built in SuperCollider, Philips Hue light scenes driven as an atmospheric channel, and a multimodal read of the room combining several of these signal channels at once. That work produced a handful of standalone tools that outlived the sprint itself and now sit in his own GitHub account:cast-mcp,nats-mcp,sclang-mcp, andosc-mcp(MCP servers wrapping each channel), plushue-listener, a standalone service piping Philips Hue state into Postgres and NATS.

The interesting part of this build wasn't any single sensor or output channel, it was holding the constraint: build something that senses and acts, but refuse the easy version of that idea that collapses into a score or a diagnosis. That boundary is what pushed the infrastructure toward ambient, ephemeral output (light, sound, media) instead of a dashboard, and it's the piece of the brief that shows up directly in the tools themselves.