Carnegie Mellon Wired a Lab Robot to Claude in 8 Hours

Carnegie Mellon wired a lab robot to Claude in eight hours

Anthropic released a research preview of the Model Hardware Standard on August 27, 2026, a shared specification that lets AI agents operate physical lab and manufacturing hardware such as microscopes, liquid handlers and robotic arms. At Carnegie Mellon, engineers connected a liquid handler, a plate reader and a robotic arm running across three separate computers to Claude in eight hours, replacing a vendor integration that typically takes several weeks.

The standard inserts a standardized driver between the operating system and the hardware, translating instrument-specific commands into simple "read" and "write" primitives that any MCP-compatible AI model can call, and it stores device information as natural-language tags an agent can interpret without a custom translator for every instrument. Anthropic said the approach cuts integration work "to hours or minutes," and Carnegie Mellon's serial dilution experiments ran about three times faster once Claude was directing the equipment, independently adjusting concentration ranges and reaching a curve fit above R-squared 0.98.

QuEra's laser recovery jumped from 58 percent to 99.3 percent

QuEra Computing, a quantum-hardware company, used the Model Hardware Standard to automate laser-lock recovery on its equipment and saw its success rate rise from 58 percent to 99.3 percent, with recovery time falling from 150 seconds to about 6 seconds on 96 percent of attempts. Anthropic said the agent rewrote a linear recovery sequence into a decision tree after iterating on the problem overnight, a change QuEra's own team had not implemented.

At the University of Washington, the Baker and Pinglay labs connected six different instruments to Claude in under a week, compared with months of failed attempts using earlier tooling, and now use the agent to watch qPCR amplification curves and halt reactions at the optimal moment. The results across three independent labs point in the same direction: a standardized hardware interface was the lever behind the speed.

Genentech, Janelia and a dozen hardware makers signed on

HHMI Janelia Research Campus co-developed the standard with Anthropic and used it to unify seven different vendor control programs into a single interface, applying it to laser beam alignment and to real-time heart-rate and neural-activity tracking during zebrafish imaging. Genentech automated a BCA protein assay across a liquid handler, robotic arm and plate reader, with Claude autonomously tuning dispense flow rates and recovering on its own from tip-pickup failures and fluid-detection errors.

Tetsuwan Scientific tested the standard against 9,143 individual dispenses across 300 transfer types and found its precision predictions beat the manufacturer's own specifications by 12 percent, and other named partners include Amazon Web Services, Automata, Danaher, Doosan Robotics, MBF Bioscience, QIAGEN, Tecan, Universal Robots, Hugging Face and Raspberry Pi. Genentech's own researchers were blunt about the limits: even strong general-purpose models "still struggle with physical, chemical, and biological constraints," and Anthropic lists hardware without a programmable interface, and tasks needing real-world physical intuition, as cases the standard does not yet solve.

Europe's Machinery Regulation lands in January 2027

The Machinery Regulation (EU) 2023/1230 becomes mandatory across the EU on January 20, 2027, replacing the older Machinery Directive and adding explicit rules for software, connectivity and automatic behavior changes in machinery, on top of the mechanical safety rules that already applied. A robotic arm or liquid handler that an AI agent now directs through the Model Hardware Standard is exactly the kind of connected, software-updatable machinery the new regulation was written to cover, and manufacturers placing such equipment on the EU market will need to reassess their conformity assessment route well before the deadline. The UK runs a parallel UKCA marking regime for machinery outside the EU framework, so UK labs piloting the standard should check the domestic requirement separately rather than assume the EU rules cover them.

The EU AI Act adds a second layer: under Article 6, an AI system counts as high-risk if it acts as a safety component of a product that already needs third-party conformity assessment under Annex I legislation, which includes the Machinery Regulation, and if a failure in that AI system could compromise the safety of the machine or the people near it. Whether an agent directing a lab robotic arm meets that safety-component test has not been tested by regulators or courts, which leaves European labs and manufacturers piloting the standard without a clear answer on who is liable if an AI-directed arm causes physical damage or injury.

What European labs and manufacturers should ask before piloting

Anthropic said it plans to open-source the Model Hardware Standard once its physical-safety evaluations and best-practice guidance with partners are further along, meaning any AI vendor's agent, not only Claude, could eventually plug into the same interface across microscopes, liquid handlers and robotic arms. That model-agnostic design is what makes this closer to an industry infrastructure move, comparable to a shared connector standard for lab robots, than a single-vendor product launch.

European labs and manufacturers evaluating a pilot should ask now, not after an incident, who carries liability when an AI-directed arm damages equipment or injures someone, whether their machinery already falls under Annex I of the AI Act through the Machinery Regulation, and what human-oversight and stop conditions the deployment needs to satisfy Article 6 if a notified body ever asks. Getting that answer before January 2027 is cheaper than getting it during a conformity assessment.

Read next: The AI Store Manager That Fired a Human  |  Anthropic Posts $11.5B Quarter, Its First Profit