TL;DR
The Model Hardware Standard, now open as an Anthropic research preview, is a specification that lets AI agents find and operate physical instruments — microscopes, robotic arms, liquid handlers — through a common driver. It is model-agnostic, reachable over the Model Context Protocol, and will be open-sourced once safety work is further along.
The problem it addresses
Laboratory and factory equipment rarely speaks a shared language. Each instrument arrives with its own interface, integration is bespoke, and connecting a rig can absorb weeks of specialist time. The standard collapses that to a driver exposing a handful of primitives — read a temperature, set a temperature — plus a discovery format so agents and devices can find one another across a network.
The more interesting part is what the driver carries alongside the commands. Operators write device characteristics in plain language, including things a manual holds and code does not: how heavy a robot arm is, what the safety limits are. From those tags the driver generates a reference file describing what the machine measures, where the boundaries sit and which settings are adjustable. That file is what makes an unfamiliar device usable by an agent.
What partners reported
The results cited are operational rather than headline-grabbing. At quantum computing firm QuEra, an agent-built controller restored a laser’s frequency lock without human help 99.3% of the time. Carnegie Mellon ran serial dilution experiments roughly three times faster, with one agent coordinating a plate reader, a liquid handler, an arm and cameras across three otherwise incompatible machines. Genentech automated a routine protein assay; Janelia unified a microscopy rig that had needed seven separate vendor programs.
Hardware vendors are building support in — among them Tecan, QIAGEN, Doosan Robotics and Universal Robots, with Raspberry Pi and Hugging Face’s robotics library also on the list.
The limits Anthropic states
The candid section matters more than the demos. A language model learns the physical world from text and images, so its spatial reasoning is weak in ways that need expert supervision: Genentech researchers had to teach Claude that foaming in a sample was a physical fault, not a software bug. The standard also cannot reach equipment with no programmable interface at all.
Looking forward
No UK regulator currently has a position on autonomous agents operating laboratory or manufacturing equipment, and the existing framework assumes a human operator who can be held responsible. British research institutions and contract manufacturers evaluating this will find the governance question arrives before the technology does — the NCSC’s August advice to assume agents will act unexpectedly was written about software, and reads rather differently when the agent controls a laser.