AI left the screen: now it wants to run the microscope

On Thursday, August 27, Anthropic introduced the Model Hardware Standard, or MHS. In plain terms: a common way for an AI agent to talk to physical machines. Microscopes, liquid handlers, robot arms, lasers inside a quantum computer. Several at once.

Until now, the agent in the news lived on a screen: it clicks in Chrome, reads a folder, sends an email. A lab is another world. Each instrument has its own software, its own way of switching on. Wiring two of them together often takes weeks, sometimes months, and someone who writes a bridge for that one bench. Anthropic says that with MHS, that bridge drops to hours or minutes.

The trick is not “the AI grew hands.” It is a driver — a translator — with two simple verbs: read and write. Read the temperature. Write the temperature. The device introduces itself in a standard format. The agent learns what the machine measures, what can be adjusted, and which safety limits must not be crossed. A lot of that knowledge today sits in a paper manual, or in the operator’s head. In MHS, that file can be written in ordinary language.

The work started with HHMI’s Janelia campus in the United States. A scientist there had a rig of lasers, motors, and cameras from different vendors, with no shared language. Anthropic is now opening a research preview to a first group of labs and factories, and says it will open-source the standard later. Any model can use it, not only Claude. The planned hook includes MCP, the same protocol many on-screen agent tools already speak.

The examples the company published are from the bench, not the kitchen. Genentech tested a classic protein assay, coordinating a liquid handler, an arm, and a plate reader. At Carnegie Mellon, an agent ran the same kind of bench across three computers that could not talk to each other, and the assay finished about three times faster. QuEra, which builds quantum computers from atoms, let an agent tend the laser “lock” — the exact frequency the beam must hold. The company says the agent recovers that lock on its own 99.3 percent of the time.

There is a large asterisk, and Anthropic puts it there. Claude learns the physical world from text and images. Space, weight, foam in a tube — that still needs a specialist at the shoulder. At Genentech, researchers had to teach it that foam in a sample was a physical failure, not a software bug. Hardware with no programming interface is out. The company talks about a physical-safety roadmap, and about evaluations against misuse, before it opens the code.

For a reader who only wants to know what is going on, the pointer is this. This week’s fight is not only the agent in the browser. It is the agent at the bench — still a preview, still with people in the loop, and already wearing the name of a standard.