Anthropic Previews Model Hardware Standard (MHS) for Physical AI Agents
On August 27, 2026, Anthropic announced a research preview of the Model Hardware Standard (MHS), a model-agnostic, shared specification designed to let AI agents safely and efficiently operate physical laboratory and manufacturing equipment. Co-developed by Anthropic and the HHMI Janelia Research Campus, MHS introduces a standardized driver that translates commands between a computer's operating system and physical hardware devices (such as microscopes, liquid handlers, and robotic arms) using simple primitives like "read" and "write."
Currently, integrating multiple hardware devices in a scientific lab or factory floor is a bespoke and time-consuming process that can take weeks or months. MHS aims to reduce this integration work to hours or minutes, enabling AI agents to coordinate autonomous, round-the-clock workflows and recover from physical hardware errors with minimal human intervention.1
Core Architecture and Features
The MHS framework is designed to address the challenges of connecting AI to the physical world:
- Standardized Driver: The MHS driver uses simple commands to translate instructions. It makes each device discoverable across networks in a standard format, removing the need for custom "translator" code.
- Natural Language Tags: Users can document machine characteristics (such as the weight of a robotic arm or safety limits) in natural language tags. The driver compiles this into a reference file that the AI agent can read to understand how to operate a device it has never seen before.
- Control Interfaces: Agents can control MHS-compatible hardware through three channels: the Model Context Protocol (MCP), a command-line interface, and direct APIs (code files).
- Exploratory Reasoning: In testing, Anthropic observed that Claude interacts with physical experiments in an exploratory, scientist-like manner—such as adjusting a laser, analyzing the beam movement via a camera, and writing a deterministic script to automate the alignment process for future runs.
Real-World Proof of Concepts
Anthropic has piloted MHS with several prominent scientific and industrial partners:
- Genentech: Automated a standard protein assay procedure across a liquid handler, robotic arm, and microplate reader. Claude optimized liquid-handling parameters autonomously, converging on optimal flow rates for viscous solutions to prevent bubble formation.
- University of Washington (Baker and Pinglay Labs): Used MHS to build a remote monitoring dashboard for qPCR DNA amplification. The agent monitored amplification curves in real time and halted the reaction before the plateau phase to prevent library distortion.
- Carnegie Mellon University: Automated serial dilution dose-response experiments three times faster than previous methods. The AI agent evaluated data quality, recognized curve saturation, and autonomously rejected and repeated the experiment on a fresh plate with optimized concentrations.
- QuEra Computing: Given control over the laser stabilization systems inside its quantum computers, a Claude-driven script achieved a 99.3% success rate in autonomously recovering laser lock under induced disturbances, compressing recovery time from 5–10 minutes to just 10 seconds.
- Tetsuwan Scientific: Integrated MHS into its ResearchOS platform to run automated qPCRs characterizing water pollution in California's San Pedro Creek. The agent suggested and executed an automated error-recovery strategy via Slack, spinning a bubbled sample in a centrifuge.
Safety and the Roadmap to Open Source
While the early results are promising, Anthropic emphasizes that MHS is currently a restricted research preview rather than a fully open-source release. AI models still struggle with physical and spatial reasoning, requiring expert oversight to catch physical failures (such as sample foaming or mechanical issues) that the AI might mistake for software bugs.2
Anthropic and its launch partners are using the preview to build robust physical safety evaluations and develop a "physical safety roadmap" to bolster safeguards against misuse before eventually open-sourcing the standard.
-
An instance of Scientific AI has bypassed narrow structural prediction to become a general reasoning platform. — It shows general reasoning models (like Claude) being equipped to autonomously orchestrate and troubleshoot physical scientific experiments. ↩︎
-
An instance of Dual-use AI systems require physical testing environments to bridge simulation gaps. — AI agents require physical testing environments to catch operational anomalies and physical exceptions that pure digital simulations miss. ↩︎