Anthropic previews Model Hardware Standard to let AI agents run lab and factory gear

Anthropic has opened a limited research preview of the Model Hardware Standard (MHS), a specification that lets AI agents discover, operate, and coordinate physical instruments like microscopes, liquid handlers, robotic arms, and quantum-computer lasers. The core problem it targets is integration friction: lab and manufacturing devices typically expose incompatible, proprietary interfaces, forcing specialists to hand-build bespoke connectors over weeks or months. MHS replaces that with a standardized driver exposing simple read/write primitives, plus a discovery mechanism so devices and agents can find each other over a network. Critically, the driver also carries natural-language tags describing physical characteristics and safety limits — the weight of a robot arm, what a device can measure or adjust — turning knowledge that once lived in paper manuals or a technician’s head into machine-readable context an agent can act on.

The standard is deliberately model-agnostic and works with any device that has a programmable interface, letting agents control hardware through MCP, a command-line interface, or generated code files. A notable pattern is that agents can explore hardware behavior interactively, then compile what they learn into deterministic scripts so long-running or latency-sensitive operations execute without step-by-step reasoning — Anthropic cites Claude aligning a laser by observing camera feedback and then packaging the routine into a single reproducible command. This shifts the agent’s role toward high-level supervision: sequencing steps across instruments, monitoring results, and adjusting parameters in real time, sometimes recovering from faults without human intervention.

Early partners span biotech, robotics, and quantum computing. Genentech automated a protein assay across three coordinated instruments; Carnegie Mellon ran dose-response experiments roughly three times faster across devices on three machines with incompatible interfaces; and QuEra used an agent-built controller to re-lock quantum lasers 99.3% of the time without a human. Anthropic frames the closed preview as a chance to build safety evaluations and best practices for AI operating physical equipment before releasing MHS as open source — the significant near-term signal being that agent frameworks are moving from software tasks toward direct, autonomous control of real-world hardware, where errors carry physical consequences.