Anthropic Previews a Model Hardware Standard Letting AI Agents Directly Operate Lab Robots and Manufacturing Equipment
Anthropic opened a research preview of the Model Hardware Standard, letting AI agents read and write to lab and factory hardware through standardized drivers, cutting integration from weeks to hours in early tests.
Overview
Anthropic has opened a research preview of the Model Hardware Standard (MHS), which the company describes as “a shared specification for AI agents to safely operate physical devices,” giving an initial cohort of scientific research labs and advanced manufacturers a common way for AI agents to control lab and factory equipment such as robotic arms, microscopes, and liquid handlers.
What We Know
- According to Anthropic, MHS standardizes hardware drivers — the software layer between an operating system and a device — around a small set of primitive commands. Heise describes this as “a small set of fundamental operations understood by most devices, such as ‘read’ (e.g., read temperature) and ‘write’ (e.g., set temperature).”
- Devices connected through MHS become “discoverable in a uniform data format, so that they can be addressed by other devices and AI agents,” per Heise, which also reports that the standard captures device-specific safety limits, such as “the weight and reach of a robot arm, or safety limits like maximum laser power.”
- Agents can direct MHS-connected hardware through any of three interfaces: “Control runs through three mechanisms: the Model Context Protocol, a CLI, and code files,” according to MarkTechPost, a description Anthropic’s own announcement corroborates.
- Anthropic says the standard’s goal is to collapse integration timelines that traditionally required custom translator code for every device: “MHS reduces this integration work to hours or minutes,” the company writes.
- According to Anthropic, researchers at Genentech “implemented and tested MHS as a proof-of-concept for automating the BCA protein assay, a standard procedure to measure total protein concentration.”
- At Carnegie Mellon University, researchers used MHS to “run serial dilution dose-response experiments about three times faster than before,” with an AI agent orchestrating a liquid handler, according to Anthropic. The full setup — from raw equipment to a completed dilution curve — took “about eight hours, versus the several weeks a vendor-built setup typically takes,” Anthropic reports.
- At the University of Washington’s Baker and Pinglay labs, “connecting six instruments through MHS took under a week, including the time I spent writing drivers for them,” according to Anthropic’s write-up.
- Quantum computing company QuEra used MHS to give an AI agent control over part of the laser system inside its quantum machines, tasking it with recovering the laser’s “lock” — the precise frequency the lasers must hold. Before the project, manual recovery “took around 150 seconds per attempt” and “only worked about 58% of the time,” per Anthropic. After iteration, the agent-built controller reached a “99.3% success rate” across 700 blind-test trials, recovering the lock in “0.9 to 5.4 seconds” for simpler disturbances and “10 to 14 seconds” for the hardest cases, Anthropic says. Over a subsequent 19-hour run, the agent’s controller “did not lose the lock once, while the expert tuned PIDs unlocked about 1.6 times an hour,” according to the same announcement.
- MHS grew out of a collaboration between Anthropic and the HHMI Janelia Research Campus, Heise reports, and other organizations named as part of the current research preview include Amazon Web Services’ Strands Robots, Hugging Face’s LeRobot, and Raspberry Pi.
- Anthropic acknowledges limits to what the agents can currently handle on their own: when a Genentech workflow hit runtime errors caused by bubbles during mixing, “Claude’s default instinct was simply to retry the operation in the same plate well with different parameters,” the company writes, until researchers intervened to codify a fix.
What We Don’t Know
- Anthropic has not published a timeline for when MHS might move beyond the initial research-preview cohort or become generally available, though it has said it intends to eventually release the standard as open source.
- Neither Anthropic nor the outlets covering the announcement have disclosed how many labs or manufacturers are currently participating in the preview, beyond naming Genentech, Carnegie Mellon, the University of Washington, QuEra, HHMI Janelia, and Tetsuwan Scientific as early users.
- It is not yet clear how MHS will handle hardware that has no existing programming interface, a gap Anthropic’s own announcement flags as a current limitation.
Analysis
The results Anthropic has published so far — from an eight-hour Carnegie Mellon integration that would otherwise take weeks, to a laser-recovery agent at QuEra that improved from a 58% manual success rate to 99.3% — point to a narrow but concrete use case: AI agents taking over repetitive, parameter-heavy calibration and orchestration tasks in already-instrumented labs, rather than open-ended physical manipulation. Anthropic’s own acknowledgment that Claude still defaults to naive retries when it hits physical failure modes it hasn’t been shown before suggests the standard, for now, functions less as a fully autonomous lab operator and more as a faster on-ramp for connecting AI agents to hardware that specialists still need to supervise.