Anthropic previews a hardware interface for AI-operated labs and factories

Anthropic opened a research preview of the Model Hardware Standard on August 27. The announcement describes MHS as a shared interface through which AI agents can operate programmable laboratory and manufacturing equipment, including microscopes, liquid handlers, and robotic arms. The work began with HHMI Janelia Research Campus, and the preview went to an initial group of research labs and manufacturers.

MHS targets the device-integration layer

Anthropic says the standard is model-agnostic and can be reached by agent harnesses through standard protocols such as MCP. Its proposed value is a common layer for discovering devices, exchanging data, coordinating multiple instruments, and exposing operations to an agent. The company reports early integrations in science, robotics, electronics, and manufacturing, but MHS was still a preview and had not yet been open-sourced.

For engineers, the key question is not whether an agent can issue a command. It is whether the interface carries enough semantics for limits, calibration state, interlocks, idempotency, timing, provenance, and recovery. A liquid handler and a robotic arm may both expose actions, yet their safe retry behavior is fundamentally different.

Physical systems need stronger acceptance tests

Anthropic explicitly notes that Claude’s spatial and physical reasoning remains limited and requires expert oversight. It gives a lab example where physical foaming had to be distinguished from a software fault, and says devices without programmable interfaces are not yet supported.

A preview evaluation should therefore start in simulation or with nonhazardous equipment. Record every command and sensor observation, define hard envelopes outside the model, require approval for irreversible operations, and test disconnection and partial-execution cases. Treat the agent as a planner inside a safety-controlled system, not as the safety controller itself.