Anthropic previews MHS: a shared standard for AI agents to operate physical lab equipment

The Model Hardware Standard (MHS) is a shared specification for AI agents to safely operate physical devices, which Anthropic released as a research preview on August 27, 2026. MHS reduces device control to two primitives, read and write, and adds a standard discovery format so agents and instruments can find each other. Preview partners include Genentech, the Baker and Pinglay labs at the University of Washington, Carnegie Mellon University, HHMI Janelia, QuEra Computing, Tetsuwan Scientific, Hugging Face, and Raspberry Pi. At QuEra, Claude recovered a laser lock 695 times out of 700 attempts, a 99.3% success rate, and Carnegie Mellon reported wiring its instruments together in about eight hours. ASAP breaks down what MHS standardizes and what is still closed, based on the original announcement.

Two primitives are the whole control surface

MHS is a specification that compresses device control into two primitives, read and write. Getting a temperature is a read; setting one is a write. On top of that sits a discovery layer that exposes devices in a standard format so agents and instruments can find each other. Device metadata carries natural-language tags for characteristics, safety limits, and operating parameters. Agents connect over MCP, a command line interface, or code file APIs. Anthropic states that MHS is model-agnostic and works with any device that has a programmable interface.

Relocking a laser fell from five minutes to under a second

QuEra Computing's quantum hardware trial produced the sharpest numbers in the announcement. Recovering a broken laser lock takes a human at the bench 5 to 10 minutes. Claude finished the same job in 0.9 to 5.4 seconds for simpler disturbances and 10 to 14 seconds for harder ones. Across 700 trials it recovered the correct lock 695 times, a 99.3% success rate. On PID tuning, a standing configuration set by a QuEra specialist measured 15.7 mV of residual error, and Claude brought it to 1.55 mV. That is roughly a tenfold improvement on a baseline an expert had already tuned.

Integration time collapsed from weeks into a single working day

The second effect of MHS shows up in wiring time rather than accuracy. Carnegie Mellon University built a serial-dilution dose-response setup in about eight hours, versus the several weeks a vendor-built setup typically takes, and reported running the experiments roughly three times faster. The Baker and Pinglay labs at the University of Washington connected six instruments in under a week, including the time spent writing drivers. Genentech automated a BCA protein assay across a liquid handler, a robotic arm, and a plate reader, concluding that about 140 µL/s was optimal for water and 10 µL/s for BSA. Tetsuwan Scientific tested 9,143 individual dispenses across 1,508 measured conditions and predicted multi-dispense precision roughly 12% more accurately than the manufacturer's specification.

The failure paths were designed before the performance was

MHS enforces device-level safety limits inside the specification itself rather than leaving them to the agent. Actions are blocked when a plate is missing, when a plate is rotated, when the reader is busy, when a camera is disconnected, when a device is unreachable, and when an emergency stop is active. Carnegie Mellon deliberately induced all six of those conditions and confirmed each one was blocked. High-risk decisions route through human approval checkpoints. Anthropic notes that an agent accidentally using excess laser power risks bleaching a sample, which is why the limits live in the standard.

Why now: the wall software agents ran into

MHS arrives because of the surface agents can reach, not because of what agents can think. For two years the agent conversation stayed inside browsers, codebases, and internal documents, all of which end at a screen. Laboratory instruments do not. Every liquid handler and plate reader ships its own driver and its own protocol, and a human rewrites that wiring for each new combination. The contrast between eight hours and several weeks at Carnegie Mellon targets exactly that wiring cost. MHS is not an attempt to make a model smarter; it standardizes the surface an already capable model can touch. The cleanest reading is that it does for instruments what MCP did for data and tools.

The denominator matters more than the 99.3%

What deserves attention in these results is the trial count and the shape of the tasks, not the headline percentage. Numbers like 700 trials, 9,143 dispenses, and 1,508 conditions come from repetition, and the five failures stayed in the denominator rather than disappearing from the writeup. The nature of those tasks matters just as much. Relocking a laser, tuning a PID loop, and running a serial dilution all have a single defined target that a machine can score. Designing an experiment, or suspecting that a clean-looking result is wrong, appears nowhere in the table. Anthropic writes that Claude struggles with physical, chemical, and biological constraints and lacks the real-world intuition needed for troubleshooting, which is why expert oversight remains part of the loop. The narrowness of the winning band is both the strength and the limit of this preview.

What the vendor list signals for buyers

The most actionable part of the announcement is the back half of the partner list, where the instrument vendors sit. AWS through Strands Robots, Automata with its LINQ platform, Danaher, Doosan Robotics, MBF Bioscience with ScanImage, QIAGEN with QIAsymphony Connect, Tecan with Fluent liquid handling, and Universal Robots are all named as adding MHS support. For a lab weighing adoption, the first question is not whether the standard is good but what share of its existing instruments expose a programmable interface at all, because equipment without one falls outside the current scope. Raspberry Pi committing to MHS across its product line points the other way, toward small labs and workshops that never had six-figure automation to begin with. Doosan Robotics appearing in the first cohort also gives Korean industrial robotics an early seat rather than a late port.

Still closed: open source later, waitlist now

MHS is not a standard anyone can download today; it is a research preview limited to a first group of selected research labs and advanced manufacturers. Anthropic says it is sharing an early version with partners ahead of making the standard open source, and no date is attached to that step. Access runs through a waitlist at modelhardwarestandard.com. While the specification stays closed, no outside party can check how compatible the vendor implementations actually are with one another. A structural question also remains, since the same company owns both the standard and the flagship model running on top of it. Anthropic explicitly calls MHS model-agnostic, so the real test of that claim arrives when a different model drives the same instruments through the same spec.

Source: Anthropic, "Previewing the Model Hardware Standard" research preview announcement (August 27, 2026; read and write primitives, standard device discovery, MCP, CLI and code file API connectivity, model-agnostic design, QuEra 695 of 700 relocks at 99.3%, 0.9 to 14 seconds, 15.7 mV to 1.55 mV, Carnegie Mellon about eight hours and roughly three times faster with six blocked safety conditions, University of Washington six instruments in under a week, Genentech 140 µL/s for water and 10 µL/s for BSA, Tetsuwan 9,143 dispenses across 1,508 conditions and roughly 12%, partner and vendor lists, planned open sourcing and waitlist, stated limits on physical, chemical and biological reasoning), compiled by ASAP.