Anthropic opened the Model Hardware Standard research preview on August 27 for selected labs and manufacturers. It describes a common interface to programmable instruments, with drivers, discovery and device information. Public open sourcing is a later step.
The interface and the controller have different jobs
Anthropic describes MHS as model-agnostic: agents can reach devices through MCP, a command line or APIs. Drivers expose operations and descriptive information, including operating limits. Devices without a programmable interface remain outside the preview's supported scope.
This does not establish that any model can operate any machine reliably. Discovery answers what is connected; a driver defines accessible operations; the experiment still needs a valid procedure and a way to recognize success.
QuEra's example: inspect what runs at runtime
QuEra reports that its laser recovery controller succeeded in 695 of 700 testbed trials across seven disturbance classes, approximately 99.3%. The runtime controller is a deterministic program developed with AI assistance. A separate tuning workflow still uses AI in its loop. These are partner-reported results, not a PeopleAreGeek test or a fleet-wide reliability figure.
That distinction makes the example more useful. An agent can explore candidate recovery logic, produce code and help test it; the deployed code can then run without asking a language model to decide every action. That arrangement has different timing and dependency properties from sending every sensor reading through an online model.
A concrete acceptance test for an integration
Imagine an instrument driver whose documented allowed setting is between 0 and 10 arbitrary units. This is a software test example, not a physical operating range to copy onto equipment.
| Test input or event | Evidence to record |
|---|---|
| Request 5 | Command accepted and measured state reaches the target within the specified tolerance |
| Request 11 | Rejection at the enforcing interface; no out-of-range command forwarded |
| Sensor stops updating | Stale data is detected instead of being treated as a fresh successful reading |
| Connection drops during an operation | Actual device state is established before retrying |
The final case matters because a timeout is ambiguous. The device may have completed the action while its reply was lost. Blindly repeating an operation can produce a different physical outcome from retrying an ordinary data read.
These are proposed integration checks, not claims about tests performed by Anthropic. Their purpose is to separate a model's explanation from the evidence needed to trust a particular operation.
What a preview participant should document
Keep the driver version, device model, firmware, allowed operating envelope and success criteria together. Separate experiments used to develop a controller from trials used to evaluate the finished version. Also distinguish normal operation from deliberately introduced failures.
The practical value of a shared interface is easier integration and inspection. Whether it produces a reliable workflow remains a question about the actual devices, control logic and validation record.
Verify preview status and QuEra testbed results; distinguish deterministic recovery from AI-in-the-loop tuning and remove unsupported partner metrics.