The harness for physical AI

Every machine describes itself at runtime, so agents see exactly what’s connected: hardware, modules, and services. Configuration is declared, versioned, and reversible.

Arm, camera, and gripper APIs are the same across vendors, so agents write a skill once and it runs on any machine with those components. Each version rolls out like software: one machine first, then the fleet.

Agents propose moves. Limits live in the layer below them, not in their code. Every move is planned against declared obstacles first, and unsafe paths are refused. If the connection drops, the arm stops.

Goals, preferences, and dead ends are captured on every run, so agents don’t repeat them. What one robot learns carries to others, without exposing the underlying data.

With your approval, a skill goes into the Viam Registry, next to drivers, vision models, and other skills. Another team’s agent installs it on a different vendor’s arm, and it runs unchanged.