Until this week, AI agents lived on screens. On August 27, 2026, Anthropic announced the first research preview of the Model Hardware Standard, or MHS: a shared specification that lets AI agents safely discover, operate, and troubleshoot real physical equipment, from laboratory microscopes to robotic arms on manufacturing lines. In early pilots, Claude agents ran a drug-discovery experiment at Genentech, compressed a weeks-long imaging experiment into a day at HHMI’s Janelia Research Campus, and pushed laser stabilization on QuEra’s quantum computers from a 58 percent success rate to 99.3 percent.
If the Model Context Protocol standardized how AI talks to software, MHS is the same idea aimed at the physical world. Here is what the standard does, what the pilots show, where it falls short, and why it matters beyond the lab.
What Is the Model Hardware Standard?
MHS is a specification for connecting AI agents to devices with programmable interfaces. Anthropic developed it together with HHMI Janelia Research Campus, and the official announcement frames it as useful for any domain that runs equipment through software.
The problem it solves
Today, connecting an AI system to a piece of hardware means days or weeks of bespoke integration work, and every lab does it differently. A single microscopy rig at Janelia was controlled by seven separate vendor programs written in MATLAB, Python, and C#, none of which talked to each other. There was also no standard way to make that connection safe. MHS cuts integration from weeks to hours or minutes and bakes safety limits into the connection itself.
How it works
The core of MHS is a standardized driver that sits between the operating system and the device, exposing it through simple read and write primitives: read a temperature, write a new target temperature. Devices become discoverable over the network in a standard format, so agents and hardware can find each other without custom middleware. Each driver also carries natural-language tags describing what the machine is, what it measures, which parameters can be adjusted, and which safety limits the driver enforces. From those tags, the system generates reference files that give an agent working knowledge of a device before it ever touches it. Agents reach all of this through three routes: the Model Context Protocol, a command line interface, or code and APIs for orchestrating many devices at once.
What AI Agents Did in Early Testing
The pilot stories are the most concrete demonstrations yet of agents doing real physical work.
Genentech: running a drug-discovery assay
At Genentech, Claude coordinated three instruments, a liquid handler, a robotic arm, and a plate reader, to run a standard protein assay. The agent’s first attempt applied one generic liquid-handling setting to both water and a viscous protein solution, producing inaccurate transfers. Asked to sweep a range of flow rates and compare its results against an expert-prepared reference, it converged on parameters automation specialists confirmed were sound: roughly 140 microliters per second for water and a gentle 10 for the protein solution, recovering from several equipment errors along the way. That calibration work normally requires a specialist writing custom code for every parameter set.
HHMI Janelia: seven programs become one dashboard
At Janelia, MHS unified that seven-program microscopy stack into a single interface. Adding a new camera used to be a multi-day integration project; with an MHS driver, it took minutes. Launching an experiment dropped from seven failure-prone manual steps to one click, and a full imaging experiment that used to take weeks compressed into a day. Researchers also gained something simpler: instruments now report status to one dashboard, so a scientist can watch the whole lab from a laptop. The setup now supports remote monitoring of qPCR runs with automated curve analysis, and closed-loop microscopy in which the agent adjusts imaging parameters in real time while the experiment is still running.
QuEra: the overnight agent loop that beat the experts
The QuEra result is the flagship. Quantum computers depend on lasers held to extraordinary frequency precision, and when a laser loses lock, recovery took an expert 5 to 10 minutes with a 58 percent success rate. QuEra and Anthropic set up an overnight loop of four Claude instances: one proposed hypotheses, one rewrote the recovery scripts, one executed them against live lasers, and one analyzed the logs and recommended the next change.
By morning, recovery took 6 seconds instead of 150, and success rates hit 96 percent, later confirmed at 99.3 percent in blind testing: 695 successful recoveries in 700 trials. One detail deserves emphasis: the finished recovery controller is ordinary, inspectable software. The agent’s job was discovering and refining the procedure; production runs it as deterministic code, with no AI in the loop. Claude had restructured the linear recovery script into a decision tree keyed to live instrument readings. In a follow-up, the agent spent 16 unattended hours tuning 12 interdependent controller parameters, cutting residual noise roughly tenfold, and its tune held a 19-hour continuous lock while the expert-tuned baseline dropped out more than once an hour. When independent specialists audited the result, the agent’s parameters matched expert settings almost everywhere, with one exception in the agent’s favor: at a particular resonance frequency, it had removed roughly a thousand times more noise than manual tuning ever had, because the human method relied on a rule of thumb that missed it.
The Honest Limits
Anthropic is unusually direct about what did not work. During the Genentech pilot, air bubbles formed while mixing a solution, and Claude responded by retrying the same operation in the same well, creating more bubbles. A human technician would have understood the physics immediately and mixed more gently. The lesson, in the company’s own framing: large language models learn about the physical world through text and images, so their physical intuition has real gaps, and expert oversight is still required. That is the stated reason MHS launches as a research preview: Anthropic wants more safety evaluations before open-sourcing the standard.
The Safety Layer
Safety in MHS is enforced at the driver level, below the agent, so a model cannot talk its way past a hard limit. In testing, six failure scenarios, including a missing sample plate, a disconnected camera, and an active emergency stop, all correctly blocked operation before any movement. Anthropic says it is building a physical safety roadmap and will publish findings from the preview as deployment guidance when MHS is open-sourced.
What the Skeptics Say
The replies under the announcement were not all applause. The sharpest criticism argued that safety standards should have existed before autonomous agents shipped, not two years after, and the jokes about robot uprisings carried a real question underneath: what stops a bad actor from pointing agent-operated hardware at harmful goals? Anthropic’s answer so far is the driver-level enforcement plus a physical safety roadmap it says the preview exists to build. A second fair critique is coverage: MHS only works with equipment that already has a programmable interface, which leaves a great deal of older lab and factory hardware outside the standard until manufacturers add drivers, and drivers themselves still have to be written and validated per device. It is also worth remembering that every published example is a partner proof of concept rather than an independent evaluation, and several demos paused mid-run to wait for human approval.
Who Is Already on Board
The launch comes with a substantial partner list, a signal this is meant as an industry standard, not a Claude feature.
| Partner | What they are doing with MHS |
|---|---|
| Amazon Web Services | Support via Strands Robots, its agent-to-device library |
| Tecan | Adding MHS to Fluent liquid handling platforms |
| Universal Robots | Planned support across its robotic arm platform |
| QIAGEN | Proof of concept on its nucleic acid purification system |
| Doosan Robotics | Testing robotic arms for automated quality assurance |
| Automata | Adding MHS to its LINQ lab automation platform |
| Danaher | Exploring smart instruments and autonomous labs |
| MBF Bioscience | Building a driver for laser-scanning microscope software |
| Hugging Face | Early adoption in the LeRobot robotics library |
| Raspberry Pi | Camera driver integration |
The last two matter for hobbyists and developers: plenty of people already use Claude Code to drive boards and cameras, and Anthropic says the preview will help extend MHS to consumer-grade devices so everything works under one interface.
Why This Matters Beyond the Lab
The strategic pattern is familiar. Anthropic released the Model Context Protocol in late 2024, and it became the de facto standard for connecting AI to software tools across the industry. MHS is the same play for the physical world, launched with heavyweight partners and an open-source commitment. If it lands the same way, the standard interface between AI and machines, in laboratories, factories, and eventually homes, could be set by this specification.
The near-term effect is speed of science. Agents that monitor instruments around the clock, recover from errors, and tune parameters overnight remove the slowest component from many experiments: waiting for a human to be available. At the University of Washington, a researcher connected six instruments in under a week, drivers included, and the resulting setup coordinated a robotic arm with a liquid handler precisely enough that the arm waited for dispensing to finish before moving a plate. The researcher’s own summary of the change: time once spent watching instrument curves now goes into planning experiments and analyzing data.
How to Join the Research Preview
Anthropic is inviting scientific labs, advanced manufacturers, hardware vendors, and the software companies that support them to apply through the MHS page on its site. Participants get early access and a voice in shaping the standard. For everyone else, the practical move is watching which equipment vendors announce MHS drivers, because that list defines where agent-run automation arrives first.
What This Means for Your Skills
The deeper signal in this launch is about work. Agents are crossing from screens into equipment, and the people thriving in the QuEra and Janelia stories are not the ones doing the manual tuning anymore; they are the ones who know how to direct, verify, and supervise AI doing it. That supervisory skill set, writing clear instructions, judging outputs, knowing where the model’s blind spots are, is learnable now, before your industry’s version of MHS arrives. Coursiv teaches exactly these foundations through short, structured daily lessons on practical AI skills, designed for busy people without a technical background, from effective prompting to understanding what AI can and cannot yet do. Check the official site for current course details and pricing.
What to Watch Next
Three markers will show whether MHS becomes the next MCP. First, the open-source release and how quickly rival labs and robotics companies adopt or counter it. Second, whether the safety evaluations from the preview become public deployment guidance, which regulated industries will need. Third, the extension to consumer hardware: the moment MHS drivers ship for the boards, cameras, and printers hobbyists already own, agent-operated hardware stops being a laboratory story.
FAQ
Can anyone use MHS today?
Does MHS mean AI can run experiments without humans?
Is this only for Claude?
The bottom line: MHS is the clearest sign yet that the agent era will not stop at your browser. The pilot numbers, a 99.3 percent success rate where experts managed 58, are striking, but the bubbles story is just as important: AI in the physical world is powerful, fallible, and in need of skilled human supervision. Learning to be that supervisor is the opportunity.