Anthropic opened a research preview of the Model Hardware Standard (MHS) on 27 August 2026, describing it as a shared, model-agnostic specification for AI agents operating programmable physical devices. The initial preview is available to a group of scientific research laboratories and advanced manufacturers, with examples involving microscopes, liquid handlers, robotic arms and quantum-computing hardware.

The standard is intended to address a practical problem in physical automation: laboratory and manufacturing equipment often has its own software, control methods and vendor-specific integrations. MHS provides a common layer through which an AI agent can discover devices, read their state, issue commands and coordinate several instruments.

Anthropic says the preview has already been used in demonstrations involving protein assays, microscopy, dose-response experiments and laser-lock recovery. These examples are company- and partner-described proofs of concept, not evidence that general-purpose autonomous laboratories or factories are ready for unattended operation.

Contents

What the Model Hardware Standard does

Anthropic describes MHS as a shared specification that allows AI agents to operate physical devices through a common interface. It is intended to work with any device that has a programmable interface, and the company says it is model-agnostic—that is, not restricted to one particular AI model.

The proposed use is broader than controlling one instrument at a time. An agent could, for example, operate multiple laboratory devices in parallel, coordinate a drug-discovery experiment or manage several steps involved in calibrating a laser system for a quantum computer. Those are capabilities Anthropic says MHS is designed to support; the announcement does not establish that every such workflow can already be performed reliably.

Anthropic says development began in collaboration with HHMI Janelia Research Campus. The early preview is being shared with partners while safety evaluations and operating practices are developed. The company has not given a firm date for general availability or open-source release.

How MHS connects AI agents to devices

The central component is an MHS driver. A hardware driver is software that translates between a computer or operating system and a particular device. In MHS, the driver provides a standardised layer between the operating system and the equipment.

The driver exposes simple read and write operations. Reading a temperature and setting a temperature are examples of the basic primitives described by Anthropic. Devices are presented in a standard, discoverable format so that an agent can identify what equipment is available and what operations it supports.

MHS drivers can also contain natural-language tags describing a device's characteristics, adjustable parameters and safety limits. Anthropic says the system can use those tags to generate a reference file for an AI agent. That reference file is intended to provide the agent with the information needed to interact with the device, including constraints that should shape its actions.

Anthropic says MHS can be accessed in three ways:

  • through the Model Context Protocol;
  • through a command-line interface; and
  • through code files or application programming interfaces (APIs).

An API is a defined software interface through which one program can request operations or data from another program or device. Combining these access methods is intended to allow an orchestration layer to coordinate several instruments rather than treating each device as an isolated control problem.

One reported demonstration used closed-loop automation. In a closed-loop system, measurements are taken during operation and used to adjust subsequent actions. Anthropic says Claude adjusted a laser, inspected the result through a camera and then converted the learned procedure into a deterministic script.

What the partner demonstrations showed

The reported results cover different kinds of equipment and workflows. They show how a shared control layer might reduce integration work, but they do not constitute a common controlled experiment.

Genentech: liquid-transfer optimisation

Genentech implemented MHS as a proof of concept for automating a BCA protein assay. The setup included a liquid handler, a robotic arm and a microplate reader. A liquid handler transfers measured volumes between containers, while a plate reader measures optical signals from wells in a microplate.

In the experiment, Claude optimised liquid-transfer flow rates against an expert reference. For water, the reported flow rate was approximately 140 microlitres per second, with a root mean squared error (RMSE) of 0.016. For viscous BSA, the reported flow rate was 10 microlitres per second, with an RMSE of 0.181. RMSE is an error measure in which lower values generally indicate that results are closer to the reference values being compared.

The agent also encountered tip-pickup, fluid-detection and bubble-related errors. Anthropic and Genentech report that it initially retried in the same well. Researchers then provided information about bubble physics and safer handling parameters, including moving to a clean well and reducing mixing cycles. This episode is important because it shows both the potential for iterative recovery and the need for physical knowledge and human guidance when failures do not fit a simple software error pattern.

University of Washington: monitoring and instrument coordination

A Baker and Pinglay lab demonstration at the University of Washington used MHS for remote instrument monitoring, AI-supervised qPCR and coordination between a robotic arm and a liquid handler.

The report says six instruments were connected through MHS in under a week, including the time spent writing drivers. This is a reported integration result, rather than an independently measured comparison across laboratories or vendors.

Carnegie Mellon University: dose-response experiments

Researchers at Carnegie Mellon University used MHS to coordinate a liquid handler, plate reader, robotic arm and monitoring cameras for serial-dilution dose-response experiments.

CMU reports that creating the drivers and orchestration layer took about eight hours, compared with several weeks for a typical vendor-built setup. It also reports that the system ran the experiment roughly three times faster than before.

The system blocked six artificially induced unsafe conditions before device movement:

  • a missing plate;
  • a rotated plate;
  • a busy reader;
  • a disconnected camera;
  • an unreachable device; and
  • an active emergency stop.

For the experimental data, the system rejected the initial serial-dilution run after saturation produced an R-squared value below 0.9. R-squared measures how closely data fit a model; higher values generally indicate a closer fit within the tested data. The system reran the experiment after reducing the top concentration from 200 to 100 micrograms per millilitre. The second run was accepted with an R-squared value above 0.98. The source also reports “3.4 variation” for repeated measurements but does not define that measure.

The experiment used a colourimetric dye as a stand-in for a drug candidate. Validation with real drug candidates and biological readouts remains future work, so the result does not establish performance in a drug-development workflow.

Janelia and QuEra: microscopy and quantum-computing hardware

HHMI Janelia researchers used MHS to unify and orchestrate a microscopy rig that previously required seven different vendor programs. The contribution says adding a camera to the MHS setup took a few minutes and that starting an experiment was reduced from seven separate launch steps to one dashboard action. The work remains under development and is still supervised by the researcher.

QuEra used MHS to give an AI agent control over parts of a quantum-computer laser system. It reports that the agent recovered the laser lock 99.3% of the time without human intervention. A prior bespoke recovery script succeeded about 58% of the time, according to the same contribution.

These figures describe demonstrations by Anthropic's partners. The supplied material does not provide an independent evaluation of the hardware, the test conditions or the broader set of failure modes.

Safety, supervision and remaining limitations

The source discusses device-level safety limits and planned human-approval protocols, but neither establishes that an integrated system is safe for unattended operation.

The demonstrations are research proofs of concept. They do not show that an AI agent can reliably manage arbitrary laboratory or manufacturing workflows, understand every physical and chemical constraint or recover from unfamiliar failures. Anthropic and Genentech state that current models struggle with physical, chemical and biological constraints, including troubleshooting that requires real-world physical intuition.

The Genentech example illustrates this limitation directly: the agent's initial response to bubble-related failures was not adequate until researchers supplied additional guidance. More complex protocols are expected to require significant optimisation and broader support for physical manipulation.

Other practical limits remain. Continuous agent operation can involve computing costs that must be weighed against the researcher time saved. The Janelia system still has human supervision, and it remains unclear how much supervision would be required for different hardware classes or higher-risk operations.

The supplied announcement also does not include a full MHS specification, a formal protocol definition, an implementation repository or an independent security evaluation. It does not establish how broadly the standard has been tested across device vendors, operating systems, hardware types or failure modes.

What happens next

Anthropic says it will use the research preview with partners to develop safety evaluations and best practices before making the standard open source. The direction is technically significant because it places the interoperability layer—the drivers, discovery system and orchestration interfaces—between AI models and physical equipment.

Whether that layer becomes a broadly useful standard will depend on factors not yet resolved: the completeness of the specification, the quality and security of individual device drivers, the handling of unsafe commands, the reliability of agents under unfamiliar conditions and the amount of human approval required.

For now, MHS is best understood as an early attempt to standardise how AI systems interact with programmable laboratory and manufacturing hardware, supported by partner demonstrations but not yet by evidence of general-purpose autonomous operation.

Sources