MHS Gives AI Agents a Common Interface for Lab Equipment

For complicated experiments, connecting instruments can itself take weeks of software work. Arco Bast says that if a lab wants a camera and a microscope stage to work together, it must build a custom integration between them; every additional device complicates the arrangement.

The Model Hardware Standard, or MHS, is presented as a way around that bottleneck. ? alek-kemeny says Anthropic spent the past year putting AI to work in labs and manufacturing facilities and repeatedly ran into the absence of a common way to connect a model to physical equipment. MHS is intended to supply that common layer: one standard way for an AI model to connect with and operate devices.

The architecture presented makes the contrast concrete. Before MHS, a control PC sits amid individual connections to a camera, microscope, sensor array, incubator, vacuum pump, spectrometer, oscilloscope, mass spectrometer, laser, pipette robot, thermal camera, centrifuge, robot arm, and 3D printer. The problem is not simply the number of instruments, but the need for separate software links among them.

After MHS, the diagram places the standard between an agent and the equipment. Bast says each device connects once through one interface, while the agent is given the relevant context to control operations. Devices that speak MHS can communicate with one another, he says, at “bare-metal speed.” The proposition is connective infrastructure: compatible instruments do not require a bespoke link for every new device pairing or AI-controlled workflow.

A Leica microscope can be focused autonomously, used to look for bacteria, and directed to decide what to capture next. Kemeny describes Claude doing those tasks in a Danaher deployment, alongside a live microscope feed, the physical hardware, and running code logs.

At Genentech, Kemeny says, a scientist designed an experiment in a PDF and dropped that PDF into Claude. Claude then autonomously executed the experiment and, when things went wrong, recovered—including overnight. The workflow begins with an experimental plan rather than a sequence of manual instrument commands.

For Bast, the practical change is a faster iteration cycle and a shift in the scientist’s attention toward the scientific question rather than the coordination of devices.

The HHMI Janelia Research Campus work gives a more immediate version of that interaction. While examining images of neurons in the brain, Bast describes instructing a microscope to move around, go deeper, or take side views in real time. The scientist directs the observation; the connected system carries out physical changes to the instrument as the work proceeds.

Bast describes experiments that took weeks taking days. Separately, he says MHS reduces the complexity of device integration by having each instrument connect through one interface, and that setup is accelerated by the agent.

The operating model is that, once hardware is connected through the shared interface, an agent has access to device context and can direct operations. The deployments described by Kemeny and Bast put that model into practical terms: autonomous focusing and image selection on a microscope, execution and recovery of a PDF-defined experiment, and real-time repositioning of a microscope viewing neural samples.

Kemeny extends the argument to scientific output. If researchers can test hypotheses faster, he says, they could create general technologies faster—such as new materials—while addressing important scientific problems. His broadest formulation is that this could compress “a century of progress” into a decade. The concrete cases center on the operational premise beneath that ambition: AI agents working through connected physical equipment.