Future World Forum

AI 2 min read

Anthropic has published a standard for letting AI agents operate laboratory hardware

Four small paper instrument shapes of different outlines, each joined by a single identical connector to one central band

Software agents have spent two years getting good at software. The interesting frontier is the part of the world that does not have an API: a microscope, a pump, an arm.

What was published

Anthropic has opened a research preview of the Model Hardware Standard (MHS), a specification that lets AI agents discover and operate physical instruments — microscopes, robotic arms, liquid handlers, cameras — through one common driver rather than a separate integration per vendor.

The primitives are deliberately small: read a value, write a value, plus a discovery format so that agents and devices can find each other on a network. It is model-agnostic and reachable over the Model Context Protocol, the company’s earlier standard for software tools.

What it claims to save

Connecting instruments today, in Anthropic’s account, «typically requires bespoke software that takes specialists weeks or even months to build». The company’s claim for MHS is that the same work drops to minutes.

Anthropic also notes something about behaviour rather than plumbing: «We’ve found that Claude interacts with experiments and hardware in an exploratory manner, much as a scientist would.»

The objection, from the field

Kaoutar El Maghraoui, principal research scientist at IBM, put the reservation plainly: «It is an impressive proof of concept, but how do we ensure safety in the physical world?»

MHS is currently a research preview with selected partners rather than an open-source release, which is itself an answer of sorts: the safety work is the reason the specification has not been thrown over the wall.

Standards of this kind are usually judged on elegance and decided on adoption, and adoption here depends on a party that has not spoken yet: the instrument manufacturers. A common driver interface is a threat to a vendor whose control software is part of what the customer is locked into, which is exactly the dynamic that kept laboratory automation fragmented for thirty years before anyone involved AI. The safety question is also more specific than it sounds. An agent that misreads a software tool returns a wrong answer; an agent that misdrives a liquid handler produces a physical outcome that no undo covers — and the interesting design detail in MHS is that operators describe device characteristics in plain language, including things a manual holds and code does not, such as mass and safety limits. That is a sensible place to put constraints, and a poor place to leave them unaudited.

Source: Singularity Hub checked against the source

More Dubai news

A torn periwinkle paper band across a deep indigo ground with a small cream marker standing at each end

AI

Two questions, one trial

A Nature Biotechnology paper proposes measuring biological age inside ordinary disease trials, using blood already drawn, and sets out what regulators would need first.

A row of small paper squares of the same size on a deep blue ground, most of them pale blue and two of them cream

AI

Claude's text carries a watermark now

Since 2 August every new Claude model has been marking what it writes, using a keyed pattern in its own word choices. The detector is deliberately not public.

A grid of small periwinkle paper squares on an indigo ground, a few of them lifted forward and joined by pale lilac threads

AI

Agentic AI training for 14,000 firms

A new e-learning platform, Dubai Chambers Academy, is putting agentic AI tracks in front of the Chamber of Commerce's business groups and councils.