Closing the Loop: When AI Gets Its Hands on the Lab Bench

In February, we wrote about AI's shift from diagnostic tool to research collaborator: systems such as Google's AI co-scientist that can generate novel, testable hypotheses in days rather than years. Since then, a quieter development has caught our attention, and it may matter more.

Hypotheses are cheap. Experiments are not. An AI that can propose a thousand candidate protein designs or a hundred plausible drug targets does little to accelerate discovery if every one of those ideas still has to queue for a human to pipette, incubate, image and read out the result. The design-build-test-learn cycle at the heart of life sciences R&D only speeds up if AI can reach the physical bench. Until recently, it largely couldn't.

The problem: instruments that don't talk to each other

Anyone who has worked in a research lab will recognise the situation. A typical automated workcell combines a liquid handler, a robotic arm, a plate reader and perhaps a centrifuge or thermocycler, each from a different vendor, each with its own software, data format and control interface. Wiring them together is a bespoke integration project that takes weeks or months and requires a specialist automation engineer. As a result, automation has historically only paid off for protocols repeated at enormous scale, such as high-throughput screening. The flexible, ever-changing experiments of a typical academic or early-stage biotech lab have stayed manual.

The Model Hardware Standard

On 27 August, Anthropic opened a research preview of the Model Hardware Standard (MHS), developed in collaboration with HHMI's Janelia Research Campus. The idea is simple. MHS defines a standardised driver, built around a small set of primitives such as "read" and "write", which any device with a programmable interface can expose. Each device becomes discoverable on the network in a common format, and the driver carries a plain-language description of what the instrument measures, what can be adjusted and, crucially, what safety limits will be enforced. An AI agent can then operate a device it has never encountered before, and orchestrate several at once, via standard protocols such as the Model Context Protocol. The standard is model-agnostic, works with any AI system, and is slated for open-source release once safety evaluations are complete. Anthropic's own analogy is USB-C: a standard interface rather than a proprietary cable.

Early results from life sciences labs

What makes this more than a specification is the evidence already emerging from the preview partners:

• Genentech used MHS to automate a BCA protein assay across a liquid handler, robotic arm and plate reader. The AI agent autonomously optimised pipetting flow rates for aqueous and viscous protein samples, arriving at parameters that Genentech's automation experts confirmed were reasonable, and recovered on its own from tip pickup and fluid detection errors.

• Carnegie Mellon researchers integrated four instruments spread across three computers with mutually incompatible interfaces in about eight hours, and ran dose-response experiments roughly three times faster. When the first serial dilution saturated, the agent rejected the curve, compressed the concentration range and reran the plate without human input.

• At HHMI Janelia, a microscopy rig that previously required launching seven vendor programs in a fixed order now starts with one click, and adding a new camera dropped from a multi-day job to a few minutes.

Just as significant is who is building support into their products. Tecan is adding MHS to its Fluent liquid handlers, QIAGEN has a proof-of-concept on QIAsymphony Connect, Automata is integrating it into its LINQ platform, MBF Bioscience is building a driver for ScanImage, and Danaher is exploring how MHS could scale its autonomous laboratories. Standards are won by distribution rather than elegance, and these are the instruments that already sit in life sciences labs.

Why a standard changes the picture

The AI co-scientist we described in February operates on literature and data. MHS gives that kind of system a route into the wet lab: proposing designs, running the builds and assays, reading results back through the same interface, and planning the next round. The University of Washington team describes exactly this vision for de novo protein design, and considers it reachable on an academic budget. If integration costs stop scaling with the number of instruments, the pool of labs that can run closed-loop, round-the-clock experimentation expands dramatically.

The caveats

Anthropic is candid about the limits. Large language models learn about the physical world through text and images, and their physical intuition is still poor. Genentech's team had to explain to Claude that a run of errors was caused by bubbles in a viscous sample rather than a software fault, and that retrying in the same well would only make matters worse. For now, MHS suits supervised, bounded workflows where a human defines the protocol and the agent handles execution, monitoring and recovery.

For those of us in IP, there are interesting questions ahead too. When an agent independently optimises the parameters that make an assay work, who contributed to the invention? And how should experimental data generated overnight by an autonomous lab be documented to support a later patent filing? Those are topics for another edition.

For now, the direction of travel seems clear. AI has been getting better at asking scientific questions. With a common standard for operating lab hardware, it may soon be able to answer them.