Anthropic moves AI agents out of software and into the lab

Anthropic has opened a research preview of the Model Hardware Standard , a specification that lets AI agents discover and operate physical devices – microscopes, liquid handlers, robotic arms, lasers – without a custom integration for each one.

The claim is that work that takes a laboratory weeks or months, and usually needs a specialist, drops to hours or minutes. The evidence Anthropic published alongside the announcement is more interesting than the claim.

QuEra Computing builds quantum computers using neutral atoms. Its lasers must hold their frequency to roughly one part in a trillion, and when a door opens or the temperature shifts, that lock drops. A human operator takes five to 10 minutes to recover it. In a university lab, somebody drives in at 2am to do it.

QuEra had already tried to automate this. A team of four – a laser systems engineer, a software engineer, an algorithms specialist and a tester – spent several months building a script that recovered the lock 58% of the time, taking about 150 seconds per attempt.

The company then handed the same problem to Claude through MHS. Four instances of the model ran a loop overnight: one proposing changes, one writing them, one running them against the live laser, one reading the logs and deciding what to try next. By morning, recovery was taking about six seconds. In a later blind test across 700 trials, the script it had written recovered the lock 695 times – a 99.3% success rate .

QuEra then pointed it at lock quality, governed by 12 interdependent parameters. Its own specialist’s tuning measured 15.7mV of residual error. Over 363 experiments and 16 unattended hours, the agent got it to 1.55mV. The specialist retuned the laser from scratch by his usual method, without seeing the result, and both were measured on a phase noise analyser. They matched across the band except at one resonance, where the manual tune had left about a thousand times more noise. Over a 19-hour run, the agent’s settings never lost the lock. The expert’s dropped it about 1.6 times an hour.

MHS introduces a standardised driver with a small set of commands – read a temperature, set a temperature – that any device can act on, and makes each device discoverable across a network. It also lets users describe a machine’s characteristics in plain language, including things not discernible from code, such as how heavy a robot arm is and therefore how carefully it must be moved.

Carnegie Mellon researchers wired up a liquid handler, a plate reader, a robotic arm and monitoring cameras spread across three computers with incompatible interfaces – one of which had no programmatic interface at all, only a screen – in about eight hours, against the several weeks a vendor-built setup takes. A University of Washington doctoral student connected six instruments in under a week, including writing the drivers.

Claude learns about the physical world through text and images, it says, so its spatial and physical reasoning still needs expert oversight, Anthropic said.

A clear illustration of this comes from Genentech, which used MHS to automate a protein assay.