On August 27, 2026, Anthropic opened a research preview of the Anthropic Model Hardware Standard, a specification that lets AI agents discover and operate physical lab instruments without months of bespoke integration code. The rollout included published results from five partner labs and manufacturers, including a reported 99.3% lock retention from a QuEra neutral-atom laser system controlled entirely by an AI agent. The Anthropic Model Hardware Standard lab automation effort traces back to work begun two years earlier by Arco Bast, a postdoctoral scientist at the Howard Hughes Medical Institute’s Janelia Research Campus in Virginia.
Anthropic Model Hardware Standard lab automation debuts in research preview
Standardization in lab automation has been a stated goal for nearly two decades. The SiLA consortium, founded in 2008, worked for eleven years before releasing version 2.0 in 2019. As of 2026, SiLA still lacks a universally adopted open communication standard, and manufacturer participation remains uneven. The gap is structural: SiLA standardized how devices are commanded, defining verbs such as “aspirate” for a pump, but left each device’s physical characteristics weight, range of motion, safe operating speeds, measurement capabilities and parameter limits in paper manuals or engineer-authored driver code.
When Bast set out to image thousands of neurons simultaneously across a rig assembled from seven vendors, the integration problem alone threatened to consume the experiment. The solution he developed at Janelia became the technical foundation Anthropic later productized. Bast’s team published the shared-memory architecture in 2024, and Anthropic hired him to lead the hardware standards program. The Anthropic Model Hardware Standard lab automation preview formalizes that work as an open specification available to any manufacturer or lab.
Prior lab automation standards standardized commands but not device knowledge
The distinction matters because AI agents need structured information about a device before they can safely operate it. A driver that only accepts a “move arm to position X” command tells an agent nothing about whether that position lies outside the arm’s mechanical envelope, whether the attached plate weighs more than the rated payload, or whether moving at default speed will vibrate a sensitive microscope stage. Anthropic’s MHS driver specification addresses this through structured natural-language fields called tags. A manufacturer or operator fills in the device’s weight, measurable quantities, adjustable parameters and safety limits. The driver then compiles those tags into a machine-readable reference file that an agent can consume on first contact.
Standardizing device knowledge rather than only device commands is what makes MHS architecturally different from SiLA and prior efforts, and what allows it to function as an AI-native interface. Anthropic positions the Anthropic Model Hardware Standard lab automation framework as a complement to existing standards rather than a replacement, but the specification’s structured-knowledge layer has no direct precedent in the SiLA ecosystem or in proprietary vendor APIs.
How MHS drivers, shared memory and three-channel control work
The Anthropic Model Hardware Standard lab automation specification introduces three components that together connect language-model reasoning to physical hardware. The first is a standardized driver that exposes a uniform set of primitives such as “read” (for example, “get temperature”) and “write” (for example, “set flow rate”). The driver also publishes each device on the network in a standard discovery format, so agents can locate available instruments without a custom translator per connection.
The second component is a shared memory dictionary, inherited directly from Bast’s Janelia architecture. Instead of each instrument keeping its state in private, OS-gated memory, every connected device writes its current state to a common memory pool that any other device or agent can read simultaneously. In the QuEra deployment, the scanning mirrors’ position, the stage location and the laser gating signals all update in the same memory region, allowing the laser to gate in step with the mirrors and the stage to compensate for sample drift without a central orchestration process.
The third component is a three-channel control interface. Agents can reach MHS devices through the Model Context Protocol, through a command line, or through pre-compiled code files. The compiled-code channel addresses latency: when a task runs faster than an agent can reason in real time, the agent learns the sequence through repeated trials, compiles it into a deterministic script such as an inspectable Python file, and runs that script without per-step inference. The QuEra laser-lock controller operates this way after the agent learned the recovery procedure across hundreds of overnight trials and converted the policy into standalone code that maintains lock without an agent in the loop.
Eight hours of upfront work to teach an AI a new instrument
For researchers who have never connected an AI to hardware, the most consequential question is how much work the Anthropic Model Hardware Standard lab automation specification requires upfront. Carnegie Mellon University researchers built MHS drivers for four instruments: a CyBio Felix liquid handler, a Varioskan LUX plate reader, a Thermo Scientific Spinnaker robotic arm, and a set of monitoring cameras distributed across three computers. The total time reported was approximately eight hours of fundamental engineering work, with the bulk spent adapting the Spinnaker arm’s motion-planning interface rather than writing integration glue.
Other partner labs reported similar figures. Howard Hughes Medical Institute’s Janelia Research Campus documented integrations with custom-built two-photon microscopes and behavioral arenas. The Broad Institute reported MHS compatibility with its existing sequencing pipelines. Anthropic published the specification, reference drivers and partner results simultaneously on August 27, 2026, and stated that additional manufacturer adoption will be tracked through a public registry. The Anthropic Model Hardware Standard lab automation preview is positioned as the first AI-native hardware standard in active production use across multiple instrument classes.