Anthropic wants AI agents to safely run lab equipment, and is testing a shared standard for it

Anthropic announced the Model Hardware Standard on August 27, 2026, describing it as "a shared specification for AI agents to safely operate physical devices." It targets the instruments in a research lab or on a manufacturing floor: microscopes, liquid handlers, robotic arms, plate readers, qPCR machines and laser systems. Per Anthropic, it works with "any device that has a programmable interface."

Model-agnostic, alongside MCP

MHS is not Claude-exclusive. Anthropic says it is "model-agnostic, and any agent harness can access it using standard protocols, such as the Model Context Protocol," so a Claude, GPT or Gemini agent could in principle drive the same instrument through the same interface. Anthropic is sharing an early version with a first cohort that includes Genentech, Carnegie Mellon, HHMI Janelia, QuEra, the University of Washington and Tetsuwan Scientific. Fortune separately reports additional early partners including Universal Robots, AWS, Doosan Robotics, Danaher and Hugging Face. An open-source release is planned once safety evaluations are further along.

The boundary problem, now physical

The recent OpenAI and Anthropic testing incidents covered here were failures of the boundary between an agent's reasoning and its access to a system, in software. MHS extends that question to physical actuation. A standard interface that makes it easier for an agent to drive a robotic arm or a laser also removes friction from the class of action that went wrong in software. One shared, documented interface is a better place to build safety controls than a dozen custom ones. It is only as safe as the runtime check that the agent issuing a command is authorized to send that command to that device.