What is MHS?
Anthropic has announced the Model Hardware Standard (MHS), a software specification that lets AI agents safely operate physical lab devices — with the stated goal of cutting instrument integration from "weeks to months" down to "hours to minutes." The announcement came with quantified pilot results from six institutions.
MHS is a pure software standard: it defines no voltages, connectors, or mechanical dimensions. It only solves how agents and devices communicate, in four layers:
| Module | What it does | |---|---| | Standard drivers | Translate hardware into two primitives: read (query state) and write (issue commands) | | Device discovery | Devices self-describe in a standard format so agents can find them on a network | | Metadata tags | Natural-language machine descriptions (payload limits, adjustable values, safety limits) | | Three control channels | MCP, command line, direct API calls — one line of code can orchestrate a chain of devices |
Safety limits are mandatory and enforced at the driver layer. The standard is both vendor- and model-agnostic: Claude can use it, but so can agents from other providers.
Origin story
The structure grew out of a personal hack: Janelia postdoc Arco Bast wrote a shared-memory dictionary for his own brain-imaging rig. Anthropic's Alek Kemeny saw it, and the two turned it into a standard.
Six pilots, hard numbers
Pilot institutions: Genentech, University of Washington, Carnegie Mellon, HHMI Janelia, QuEra, Tetsuwan Scientific.
Carnegie Mellon — serial dilution
- Connected a CyBio Felix liquid handler, a Varioskan LUX plate reader (no API — the agent drove its GUI), a Spinnaker robotic arm, and cameras across three incompatible computers in ~8 hours; comparable jobs are normally quoted by the week.
- Researchers injected six deliberate faults (missing plate, misaligned plate, plate reader occupied, camera disconnected, device lost, emergency stop) — all caught.
- When one run returned R² below 0.9, the agent judged it a failure, changed concentrations, and reran autonomously, recovering R² above 0.98 with no human intervention.
- For neutral-atom quantum computers: the old script recovered lock in ~150 s with 58% success; a human expert took 5–10 minutes. The agent-written re-lock controller succeeded 695/700 times with recovery in seconds.
- Over 363 experiments / 16 hours unattended, residual phase-lock error dropped from 15.7 mV to 1.55 mV. Over 19 hours it never lost tuning; the human comparison lost it ~1.6 times per hour.
- Claude orchestrated three devices and optimized pipetting itself: 140 µL/s for water, 10 µL/s for viscous BSA solution — matching expert intuition. Tip-pickup and liquid-level errors self-recovered.
- The only human intervention:面对 bubbles, it couldn't tell whether the problem was physical or software — a person had to clarify.
- Claude's spatial/physical reasoning still needs expert oversight; the QuEra agent was even over-cautious, halting overnight to await approval.
- Compute costs for long-running agents are nontrivial.
- Orchestration of complex protocols still requires tuning.
- Anthropic announcement: https://www.anthropic.com/news/model-hardware-standard-research-preview
- Application portal: https://modelhardwarestandard.com
QuEra — laser phase re-locking
Genentech — BCA protein quantification
Ecosystem
Vendors integrating: AWS, Danaher, Tecan, QIAGEN, Doosan, Universal Robots, Automata, MBF Bioscience. Early adopters: Hugging Face's LeRobot and Raspberry Pi (official MHS camera driver) — meaning a $100 single-board computer can become an agent-controllable device.
Why this matters beyond lab automation
Embodied AI's dominant narrative is giving models bodies (humanoids, end-to-end VLA, world models). MHS takes the opposite route: standardize the agent-to-device interface first, then plug into the installed base of billions of programmable devices — analogous to USB for peripherals or MCP for tool calling. If it open-sources and gets built into vendor products, "agents in the lab" shifts from bespoke integration projects to simply installing a driver.
Scientific impact may be the most direct: per University of Washington figures, a de novo protein design cycle costs ~$100 and a week of validation while the design itself costs one cent — the bottleneck is entirely wet-lab side. QuEra's case even suggests quantum computer operations may be one of AI agents' earliest real jobs.
Acknowledged limitations
Per Anthropic's own announcement:
Additionally, MHS requires devices with programmable interfaces (GUI-only works); purely mechanical devices have no path in yet — Anthropic says it's working with such vendors.
The key thing to watch
The standard's open-source pledge is its most important aspect: closed standards can't build network effects. The signal to watch within six months is whether MHS compatibility becomes a printed selling point in Tecan and Danaher product catalogs.
References: