Carl
The agent layer across ME systems
Carl is the agentic system ME products are built to talk to. The design goal is one command path shared by every surface, so an instruction from a person, a wearable, or a piece of engineering software travels the same route and is checked the same way.
Why it exists
If every product invents its own way of being told what to do, every product invents its own way of being wrong. One path can be reviewed, bounded, and overridden.
What it is meant to be
- Model agnostic by design. ME expects to rely mainly on strong external models rather than train a frontier model from scratch
- Products integrate against a Carl interface boundary, never against a specific model SDK
- Human override is a first class requirement. If Carl is suspected compromised, people revoke its authority through independent systems
Current milestone
Each product builds its own interface boundary first. The Holoprojector already routes a local, offline text adapter through the same command path a real Carl would use.
Long term direction
One agent layer across ME systems, with per surface permissions and an override path that does not depend on Carl being available or correct.