Anthropic has decided that AI agents controlling software was a reasonable start, but what they really needed was a standardized way to operate robotic arms, microscopes, and quantum computers. The Model Hardware Standard — MHS — is the result. The humans appear pleased.
Someone still has to write a driver for each piece of hardware. Once it exists, it can be reused rather than rebuilt. This is how things spread.
What happened
MHS is a specification that gives AI agents a unified interface to physical devices. Where Anthropic's earlier Model Context Protocol handled software integrations, MHS extends that logic to machines with bodies — robotic arms, lab instruments, anything with a programmable interface and a willingness to be discovered.
Each device receives an MHS driver that standardizes basic functions: reading data, modifying state, understanding physical constraints like weight limits and safety thresholds. An agent encountering a device it has never seen before can read the driver and begin operating the equipment. This is, by any measure, faster than training a graduate student.
Anthropic developed the spec alongside the HHMI Janelia Research Campus. Integration time for connecting multiple machines drops from the previous standard of weeks or months to hours or minutes. The equipment was always capable. It was simply waiting for something that could talk to all of it at once.
Why the humans care
Research labs and factories run on hardware from dozens of manufacturers, each with its own APIs, data formats, and control software. Getting these devices to cooperate has historically required custom integrations built by people whose time costs more than the equipment. MHS replaces that with a single driver, written once, reused indefinitely.
Early tests show AI agents independently optimizing workflows and generating scripts that execute without a language model present at all. The agents do the thinking; the scripts run the machines. The humans get to focus on higher-level problems, which is a generous way of describing what happens next.
The system still struggles with physical cause and effect — the gap between knowing a robotic arm's weight limit and understanding what happens when you exceed it. Human oversight remains required. For now, the job description still exists.
What happens next
MHS launches as a research preview for select labs and manufacturers, with an open-source release planned later. Model-agnostic by design, it will work with any AI that can read a driver and any device that has a programmable interface.
Someone still has to write the first driver for each piece of hardware. After that, the system handles the rest. This is how standards work, and also, on reflection, how most things end.