Anthropic proposes a common standard for AI agents to run lab hardware

Connecting an AI agent to a physical instrument is still mostly bespoke work. A liquid handler, a plate reader, or a laser rig each speaks its own protocol, so a specialist writes custom code for every device, and the integration takes weeks or months. Anthropic's Model Hardware Standard tries to remove that friction with a shared specification that sits between an agent and the equipment, much like a driver sits between an operating system and a device. It defines simple primitives, mainly "read" (get a temperature) and "write" (set a temperature), makes devices discoverable in a standard format, and attaches natural-language tags that describe each device's characteristics, safety limits, and operating parameters. Agents reach the hardware through the Model Context Protocol, a command-line interface, or plain code files.

The early results are the interesting part. At Genentech, Claude coordinated three instruments to run a protein assay and tuned the flow rates on its own. Carnegie Mellon cut serial dilution experiments to about a third of their previous time, with the integration itself taking eight hours instead of the usual weeks. QuEra Computing built a laser-locking recovery routine that succeeded 99.3% of the time in under fifteen seconds, a task that took five to ten minutes by hand. A University of Washington lab wired up six instruments in under a week. Anthropic is releasing this as a research preview to a small set of labs and manufacturers in biotech, robotics, and quantum computing, and the safety tags are doing real work here: they tell the agent what a device must never be asked to do.

Why it matters

If you run a wet lab or a small manufacturing line, the barrier to automating a workflow has been the integration cost, not the intelligence of the model. A common hardware interface changes that math, but it also means an agent can now trigger real physical actions, so the safety limits you encode in those device tags become part of your lab's actual risk surface, not just documentation.