Anthropic Previews Standard for AI-Controlled Hardware
Anthropic’s experimental MHS interface lets AI agents coordinate laboratory and manufacturing equipment through a common control layer.
From software tools to physical instruments
Anthropic has introduced the Model Hardware Standard, or MHS, as a research preview for connecting AI agents to scientific and industrial equipment. The proposed interface gives an agent a consistent way to discover device capabilities, issue commands and coordinate several instruments, instead of requiring bespoke code for every microscope, camera, robotic arm or control board.
The company says early users applied MHS to real physical workflows. At Genentech, an agent ran a drug-discovery experiment while responding to errors during execution. In another trial, researchers compressed an imaging workflow that previously required weeks into days. Anthropic also cited work involving quantum-computer laser calibration and coordinated laboratory equipment.
MHS currently targets laboratory and manufacturing systems rather than safety-critical, real-time control. Anthropic acknowledged that language models still lack reliable physical intuition and said more work is required before the specification can be open-sourced. For now, scientists, robotics developers, electronics companies and manufacturers are being invited into the preview to test the design and help define its safeguards.
Why it matters
Software agents have gained useful standards for calling digital tools, but physical equipment remains fragmented across vendors, protocols and custom control systems. A broadly adopted hardware interface could give agent developers a reusable path from planning an experiment to executing it, observing results and recovering from routine failures.
That would also raise the consequences of mistakes. An incorrect software action can often be rolled back; a bad command sent to laboratory or industrial machinery may damage samples, equipment or people. MHS is therefore important less as a finished protocol than as an attempt to establish the permissions, observability and safety boundaries needed before general-purpose agents receive wider control over the physical world.