Goal-driven AI agents that perceive a spatial scene, plan, act and learn from the human standing next to them. A field report from our agent runtime.
Agents need a body
Text agents are useful. Spatial agents are transformational — because they share a coordinate system with the human user. They can point, walk, hand things over, model the user's gaze and adjust their behaviour to what the human is doing in physical space.
Perception → Plan → Act
Our runtime gives every agent three loops running at different cadences: a 60 Hz perception loop (gaze, hands, scene graph), a 4 Hz planning loop (LLM with tool use) and a 1 Hz reflection loop that updates long-term memory.
This three-loop architecture is why our agents feel calm instead of twitchy — they react fast, but they think on a human timescale.
Safety and scope
Every spatial agent ships with a scope boundary — a literal volume of the virtual world it is allowed to operate in — and a value sheet that constrains its tool use. We log every action and provide a clear way for the user to interrupt, undo or override.