Anthropic's new standard lets AI agents run real lab equipment

Claude can now physically turn the knobs in a laboratory. Anthropic’s Model Hardware Standard gives AI agents a common interface to real scientific equipment.

What launched

On August 27 Anthropic opened the research preview of the Model Hardware Standard (MHS). The standard lets an AI agent discover a physical device, understand what it can do and operate it safely, replacing the custom integration work that has kept AI out of most lab hardware. Connecting a single device used to take weeks; MHS cuts it to hours or minutes. Carnegie Mellon connected its setup in roughly eight hours instead of the usual several weeks.

The early results

The test numbers are concrete. At quantum computing company QuEra, an agent raised laser stabilization success from 58% to 99.3% and cut recovery time from about 150 seconds to about 6. Genentech ran a protein assay with an agent handling errors in real time.

Hardware support is forming around the standard: AWS, Doosan Robotics, Universal Robots, QIAGEN, Raspberry Pi and Hugging Face are building MHS into their hardware and tools. Anthropic says it will open source the standard after safety evaluations with preview partners.

What it changes

Agents stepping out of the chat window and into physical instruments compresses experiment cycles: work that took weeks of scheduling and manual operation can run in days, with the agent adjusting conditions overnight. It also moves AI safety questions into the physical world, where a wrong action affects equipment and materials rather than text. That is the stated reason the preview is limited to selected organizations, with the open source release gated on safety work.