Anthropic's Model Hardware Standard Gives AI Agents One Driver for Lab Instruments

For the past decade, the hardest part of automating a laboratory has rarely been the science. It has been the cables, the vendor SDKs, the Windows-only control software, and the brittle glue code holding it all together. Anthropic's Model Hardware Standard — a research preview that opened August 27 and is now weeks into partner testing — is a bet that the fix looks less like a robot and more like a driver.

MHS is a shared specification that lets an AI agent discover a physical instrument, read its state, and command it, using a deliberately small vocabulary. The standardized driver at its center translates between a computer's operating system and a device using primitives as plain as read ("get temperature") and write ("set temperature"). Crucially, it also carries a natural-language layer: tags where an operator — or an agent that interviews the operator — records what never made it into code, such as the weight of a robotic arm or the laser power above which a fluorescent sample bleaches. From those tags the driver auto-generates a reference file describing what a device can measure, what can be adjusted, and what safety limits will be enforced.

The device classes are unglamorous and specific: liquid handlers, robotic arms, plate readers, thermocyclers, centrifuges, two-photon and light-sheet microscopes, femtosecond lasers. Anthropic claims MHS works with essentially any device exposing a programmable interface — and Carnegie Mellon went further, driving a Thermo Scientific Varioskan LUX plate reader that had no API at all by having the agent operate its GUI the way a person would.

The Pilots

The numbers from the research preview are the story's spine. At Carnegie Mellon, researchers ran serial-dilution dose-response experiments roughly three times faster than before, going from raw, un-automated equipment to a completed dilution curve — including one autonomous rerun — in about eight hours, against the several weeks a vendor-built integration typically takes. Their setup spanned three computers with incompatible interfaces.

The most striking result came from QuEra Computing. Recovering the precise operating frequency of a titanium-sapphire laser — accurate to roughly one part in a trillion — had taken a human five to ten minutes. A bespoke script built by a four-person team over several months worked about 58% of the time and took around 150 seconds per attempt. Running unattended overnight through hundreds of iterations, an agent using MHS drove that to about six seconds at 96%. In a later blind test across 700 trials, it recovered the correct lock 695 times — a 99.3% success rate. Tuning twelve interdependent PID parameters, it cut residual error from an expert's 15.7 mV to 1.55 mV over 363 experiments.

At Genentech, MHS coordinated a BCA protein assay across a liquid handler, a robotic arm, and a plate reader. At the University of Washington's Baker and Pinglay labs, PhD student Zihao Song wired up an entire bench: "Connecting six instruments through MHS took under a week, including the time I spent writing drivers for them." Earlier automation attempts, he wrote, had meant "weeks spent evaluating platforms, chasing vendor support, learning and building glue code between instruments, and finally giving up."

Safety is handled at the driver layer rather than left to the model's judgment. Limits are declared in the device file and enforced before actuation. CMU deliberately injected six failure conditions — missing plate, rotated plate, reader busy, disconnected camera, unreachable device, active emergency stop — and reported that the system blocked all six before any device moved. Virginie Ruetten, a scientist at HHMI Janelia — where MHS originated as a collaboration between Anthropic's Alek Kemeny and postdoc Arco Bast — noted that device-level limits mean she need not worry about an agent burning her sample with excess laser power. QuEra saw the opposite failure mode: Claude paused so often for human confirmation that experiments sometimes stalled overnight.

Why It Matters

This is the moment agents stop being software that writes software. An agent that can aspirate 40 microliters, move a plate, and relock a laser is operating in a domain where mistakes are not rolled back with a git revert — they cost reagents, animals, and weeks. Anthropic is explicit about the ceiling: Claude learns the physical world through text and images, so its spatial reasoning "still require[s] expert oversight." The pilots are less a demonstration of physical intelligence than of orchestration — the agent as a tireless conductor of instruments that already knew how to play.

The strategic read is the MCP playbook, run again. MHS is model-agnostic and reachable through three mechanisms — MCP, a command line, and code files — and Anthropic says it will open-source the spec, though it has committed to no date. Alek Kemeny, a member of Anthropic's technical staff, told Bloomberg: "What MCP did for software, MHS will do for the hardware world." Owning the interface layer is cheaper than owning the model layer and stickier than either. Jonah Cool, Anthropic's head of partnerships and deployment of science, framed the pitch as liberation from proprietary control software: "We want to avoid vendor lock-in for scientists."

The vendor list is the tell. AWS is shipping support via Strands Robots, with Tecan, QIAGEN, Danaher, Doosan Robotics, Universal Robots, Automata, MBF Bioscience, Hugging Face, and Raspberry Pi building or testing integrations. Elizabeth Kelly, Anthropic's head of beneficial deployments, was candid about the endgame: "We built this for science to sort of show the promise of AI, but there's also huge benefits here for enterprise and for industry."

What to watch: whether the open-source release actually lands, and under what license. MHS is currently an application-only preview, the opposite of how MCP won. Watch whether instruments start arriving MHS-ready out of the box, whether a competing standard emerges from Nvidia or a robotics incumbent, and what lands in the physical safety roadmap Anthropic says it is developing. A standard that makes actuation trivially easy for any model is exactly as useful to the careless as to the careful.