When an AI runs lab equipment, ‘complete’ is not enough

According to Anthropic, MHS uses standard device descriptions and basic read/write commands so an agent can discover what an instrument measures, what it can adjust and which limits apply. The company says this can reduce an integration job that once took weeks or months to hours or minutes. Its early examples include coordinating a liquid handler, robotic arm and plate reader; it also reports that a partner used the system to recover a quantum laser lock 99.3% of the time without human intervention.

Those are early partner examples, not a blanket result for every lab. Anthropic is explicit about a remaining limitation: language models learn about physical conditions through text and images, and still need expert oversight. In one reported protein-sample experiment, researchers had to teach the system that foaming was a physical failure rather than a software bug. That is a useful distinction. A connected system can know every command it sent and still miss what a person at the bench notices immediately.

If an AI agent is allowed to supervise a long run, ‘complete’ is too vague to be reassuring. A better return screen would say:

- **What ran:** the instrument, sample or batch, and the time range. - **What changed:** adjustments made during the run, including retries or parameter changes. - **What did not resolve:** a stopped step, an unusual reading, or a condition the system could not confirm. - **What to check next:** one concrete bench check, with a fresh camera frame or sensor reading where that helps.

That is not bureaucratic overhead. It saves the next person from opening three vendor programs and guessing whether a cheerful green state means finished, paused, or merely connected. It also makes the right kind of interruption possible: a specialist can spend attention on the one sample or instrument that needs it instead of rereading a whole run.

NIST’s AI Risk Management Framework is deliberately not a one-size-fits-all checklist. Its playbook groups suggested actions around governing, mapping, measuring and managing risk. That fits this moment better than a generic promise that an AI system is ‘safe.’ A lab using an agent around physical equipment needs to decide which changes can proceed, which require a person to look, and what happens when the evidence is stale or contradictory.

For a small lab, the first useful deployment may be modest: let the assistant collect readings, draft a run summary and flag a possible exception, while the scientist still approves any consequential change. The prize is not removing the person from the loop. It is making their return to the loop quick enough that they can understand the state without restarting the whole investigation.

Before enabling a new connection, ask four ordinary questions: What exact device actions can it take? What will I see when it changes a setting? What makes it stop and ask for help? When I return later, can I tell the last confirmed physical state without reading a developer log?

The Model Hardware Standard is a real move toward making fragmented lab tools work together. The more interesting test comes afterward: whether its users build a handoff that lets a tired scientist understand an experiment at a glance and make the next decision with their eyes open.