Anthropic's Model Hardware Standard Puts Agents on Lab Gear

Every few months a lab announces that an AI agent "ran an experiment." What usually happened is that a human wired the instrument, wrote the driver and watched the whole time. Anthropic's Model Hardware Standard (MHS), opened as a research preview on August 27, is an attempt to make the wiring part disappear. Whether it makes the watching part disappear is the more interesting question, and Anthropic's own write-up gives a mixed answer.

What the standard actually specifies

MHS defines standardized drivers built on two primitives, read and write, that any device can expose. Devices become discoverable in a common format, and each driver produces natural-language tags plus reference files describing what the instrument can do and where its safety limits sit. An agent can reach a device three ways: through the Model Context Protocol (MCP), a command-line interface, or plain code files. Anthropic stresses the standard is model-agnostic; any agent harness that speaks MCP can use it.

The work started with HHMI's Janelia Research Campus. Preview partners are Genentech, the Baker and Pinglay labs at the University of Washington, Carnegie Mellon, the quantum computing company QuEra, and Tetsuwan Scientific. Hardware vendors backing the spec include AWS, Automata, Danaher, Doosan Robotics, MBF Bioscience, QIAGEN, Tecan, Universal Robots, Hugging Face (LeRobot) and Raspberry Pi.

The numbers, and the one that matters most

The partner results are specific. At Carnegie Mellon, researchers built a driver and orchestration for an instrument in about eight hours, against the several weeks a vendor setup typically takes. At QuEra, a laser-recovery script written by Claude lifted the success rate from 58% to 99.3% and cut recovery time from 150 seconds to 6; The Decoder reports 695 of 700 automated runs completed without a human stepping in. At Genentech the agent tuned liquid-handling flow rates from generic defaults to material-specific values (140 µL/s for water, 10 µL/s for BSA). Tetsuwan logged 9,143 individual dispenses across 300 transfer types.

Then there is the bubble. When bubbles formed in a viscous solution at Genentech, Claude kept restarting the procedure without recognizing what was physically wrong; it found a workable fix only after a human explained the cause. Anthropic included this in its own announcement, says expert oversight is required throughout the preview, and plans to build safety evaluations before open-sourcing the standard.

So is the watching part going away?

Not yet, and the bubble is why. Writing a driver is a software problem, and software is where these models are strong; that is what shrinks weeks to hours. Diagnosing why a pump is misbehaving is a physics problem, and the agent failed it in the one case Anthropic chose to publish. In a lab that costs a day of experiments. On a filling line or a CNC cell it costs considerably more, which is worth remembering given that Doosan and Universal Robots are on the vendor list and manufacturing is named as a target.

The analogy Anthropic wants you to draw is with MCP, which replaced one-off integrations with a single connection standard for software tools. If MHS does the same for instruments, the biggest beneficiaries are integrators who currently rebuild the same plumbing for every vendor protocol. The phrase to keep attached to any agent-operated machine for the foreseeable future is "under human supervision." Applications for the preview are taken at modelhardwarestandard.com.

Sources: Anthropic, The Decoder, Fortune