Anthropic previews standard for AI agents controlling physical machines

Anthropic has opened a preview of its Model Hardware Standard (MHS), a proposed system for letting AI agents control programmable machines through shared rules.

MHS uses software drivers to translate between computer systems and individual machines. Basic commands let agents read measurements or change settings, while machine details that are usually kept in manuals or by experienced staff can be entered through natural language. MHS then creates reference material describing equipment capabilities, adjustments and safety restrictions.

The standard is designed for laboratories, factories and workplaces where instruments need to work together without custom software links. It supports microscopes, liquid handlers, robotic arms and other equipment with a programmable interface. Anthropic claims integrations that once took weeks or months could take hours or minutes when equipment supports MHS.

Anthropic said Claude was tested in physical experiments, including laser alignment using camera-based observation. The model adjusted the laser, checked the image and repeated the process while assessing each change. MHS can also combine commands from several devices to create automated sequences.

Early participants include Amazon Web Services, Automata, Danaher, Doosan Robotics, MBF Bioscience, QIAGEN, Tecan and Universal Robots. Anthropic said the preview will help establish safety checks before wider availability. The company acknowledges that agents still need expert supervision and that MHS cannot currently support equipment without a programmable interface. An eventual open-source release is planned to include test findings and guidance for implementation.