Anthropic has opened a research preview of a new Model Hardware Standard (MHS), a specification designed to allow artificial intelligence (AI) agents to safely operate and coordinate physical devices used in scientific research and manufacturing.
The AI company said MHS is being made available initially to a group of scientific research laboratories and advanced manufacturers as it works with partners to develop safety evaluations and best practices before making the standard open source.
According to the firm, MHS allows AI agents to operate multiple instruments, including microscopes, liquid handlers and robotic arms, in parallel.
The technology is designed to support tasks ranging from routine drug discovery experiments to laser calibration on quantum computers.
The development of MHS began as a collaboration between Anthropic and the HHMI Janelia Research Campus.
Anthropic says integrating hardware in laboratories and manufacturing facilities can currently take weeks or months because devices generally use different programming interfaces and require specialists to develop bespoke connections between them.
“MHS reduces this integration work to hours or minutes,” the company says.
The standard is also intended to make it easier to run autonomous experiments and workflows continuously.
According to Anthropic, AI agents can reason through individual steps, change parameters as experiments progress and, in some circumstances, recover from hardware errors without human intervention.
Standardising hardware for AI agents
At the core of MHS is a standardised driver that acts as software between a computer's operating system and a physical device.
The firm explains that the MHS driver uses a set of basic commands, or primitives, including “read” and “write”. A read command could be used to obtain a device's temperature, while a write command could set the temperature.
The system also makes devices discoverable in a standard format, allowing hardware and AI agents to find and communicate with one another across networks without requiring a bespoke translator for each device.
Anthropic says MHS is designed to work with any device that has a programmable interface and is model-agnostic. AI agents can access it through standard protocols, including the Model Context Protocol (MCP).
The standard also addresses the problem of AI agents needing to understand physical characteristics that may not be apparent from a device’s software interface.
For example, Anthropic says information such as the weight of a robotic arm can be important when determining how it can be moved safely.
It note that such information is often contained in paper manuals, stored on individual computers or exists as tacit knowledge among laboratory and engineering staff.
Under MHS, users can record this information in natural-language tags within the driver. Anthropic said users can enter the information themselves or use an AI agent to interview them about their hardware setup, it adds.
The driver then generates a reference file describing a device's characteristics, including what it can measure, what parameters can be adjusted and which safety limits are enforced.
“This file gives the agent everything it needs to know to operate the device,” Anthropic says.
Autonomous workflows
Once devices are connected and understood by an AI agent, MHS provides three mechanisms for controlling them: MCP, a command-line interface and code files or APIs.
These mechanisms allow agents to orchestrate multiple devices through a single line of code, according to Anthropic.
The agent can receive operating data from connected instruments, sequence activities, monitor results and adjust parameters as conditions change.
For lengthy or high-speed operations, commands can instead be chained together in code files, allowing hardware to perform operations without the AI needing to reason through every individual step.
Anthropic says early testing with its Claude model demonstrated this exploratory approach.
“As we’ve tested MHS, we’ve found that Claude interacts with experiments and hardware in an exploratory manner, much as a scientist would,” the company says.
In one example, it states that Claude adjusted a laser, used a camera to observe how the adjustment affected the laser beam and repeated the process while attempting to understand the relationship between its actions and the resulting changes.
Claude subsequently turned what it had learned into a deterministic code file that could align the laser without requiring the model to reason through every step, allowing the process to run as a single command.
Anthropic says early projects involving biotech, robotics, quantum computing and other fields showed that MHS could reduce hardware integration times, accelerate experimentation and assist with live machine operation and fault detection.
The company is now seeking broader participation from scientific researchers, robotics companies, electronics firms and manufacturers as it develops the standard.
Anthropic points out that the research-preview phase is intended to help establish safety evaluations and operating practices for AI systems controlling physical equipment before MHS is released as an open-source standard.
Share
