When an AI agent touches a lab instrument, the tool call becomes physical

Many software-agent demos end with a file or a message. A laboratory agent might instead control a pump, microscope, heater or industrial machine. The result can be physically irreversible, and the equipment may be shared with other researchers.

That makes the Model Hardware Standard research preview^{[1]Model Hardware Standard research preview} interesting. It is being developed to describe how AI agents can interact with physical devices in scientific research and manufacturing. Its preview is not a general claim that an arbitrary model can safely control any machine.

A laboratory instrument has limits, calibration requirements, operating states and emergency procedures. A task may be safe only when a particular sensor reading is current and an operator has prepared the apparatus.

The application should know the distinction between requesting a measurement, preparing an experiment and executing an action with physical effects. The instructions that a device accepts do not by themselves provide authority to issue them.

Anthropic's Claude Science workbench^{[2]Anthropic's Claude Science workbench} is another example of attention shifting from scientific conversation to tools, data and computing resources. Our Bonfÿre work has explored the parallel requirements: account ownership, instrument descriptions, dependencies, operator limits and durable records of what happened.

A useful scientific workflow could prepare an experimental plan, check resource availability, ask an authorized human to approve a risky step and collect instrument outputs with their timestamps and calibrations.

There is no need to romanticize fully autonomous laboratories. The near-term opportunity is making scientists' existing tools work together while preserving the physical safety rules that made those tools usable in the first place.