Anthropic’s new framework will let AI agents control hardware

Anthropic has launched a research preview of a new Model Hardware Standard (MHS), a specification aimed at enabling AI agents to safely operate and coordinate physical devices utilized in scientific research and manufacturing. Initially made available to select research laboratories and advanced manufacturers, Anthropic plans to refine MHS with partners before making it open source.

The technology empowers AI agents to manage multiple instruments, such as microscopes, liquid handlers, and robotic arms, simultaneously. Capable of tasks ranging from routine drug discovery experiments to laser calibration on quantum computers, MHS promises to streamline hardware integration from weeks or months to mere hours or minutes.

This standardization simplifies the creation of autonomous experiments and workflows, allowing AI agents to reason through steps, adapt parameters during experiments, and recover from hardware errors without human intervention. Central to MHS is a standardized driver functioning as software between a computer's operating system and the physical device.

This driver employs basic commands, or primitives, like "read" and "write", enabling devices to be discoverable in a standard format, facilitating communication across networks without requiring bespoke translators for each device. Designed to work with any programmable device and model-agnostic, MHS integrates devices via standard protocols such as the Model Context Protocol (MCP).

It also addresses the challenge of AI agents needing to comprehend physical characteristics not evident from a device’s software interface, such as a robotic arm's weight, which is crucial for safe movement. Users can record this information in natural-language tags within the driver, either manually or through an AI agent conducting interviews.

The driver subsequently generates a reference file detailing the device's characteristics, enabling the AI agent to operate the device seamlessly. Once devices are linked and understood by the AI agent, MHS offers three control mechanisms: MCP, a command-line interface, and code files or APIs, allowing agents to orchestrate multiple devices through a single line of code.

These mechanisms permit agents to coordinate devices using a single command, receive data, sequence activities, monitor results, and adjust parameters as conditions evolve. Early testing with Anthropic's Claude model revealed that the AI interacts with experiments and hardware in an exploratory manner akin to a scientist. For instance, Claude adjusted a laser, used a camera to observe the effect of adjustments on the laser beam, repeated the process, and turned its findings into a deterministic code file that could align the laser without necessitating further reasoning by the model.

Early projects in biotech, robotics, quantum computing, and other fields demonstrated that MHS could reduce hardware integration times, expedite experimentation, and assist with live machine operation and fault detection. Anthropic is now inviting broader participation from scientific researchers, robotics companies, electronics firms, and manufacturers as it develops the standard, emphasizing that this research-preview phase aims to establish safety evaluations and operating practices for AI systems controlling physical equipment before MHS becomes an open source standard.