Anthropic's MHS preview keeps the per-instrument integration work inside the lab

The pitch for a driver layer is arithmetic. Right now each agent-and-instrument pair is its own engineering exercise, which is exactly how the announcement coverage describes the status quo ^[8]. A shared specification is meant to turn that product into a sum: the device publishes its capabilities once, and any conforming agent discovers them ^[3].

The pilots do not yet demonstrate the sum. Neither one describes an instrument arriving MHS-ready from its vendor. The orchestration layer at Janelia and the assay-specific implementation at Genentech were both built around the standard by the people running the rigs ^[13][11]. That is not a criticism of the design; it is where a spec always starts. But the integration count only falls when the device side of the adapter is written once by whoever ships the device, and no source here claims that has happened.

The relock figure is the most specific engineering result in the preview. Anthropic reports 99.3% autonomous laser relock recovery on a QuEra neutral-atom system, against 58% in the earlier testing description ^[7]. Read it as residual failure and it looks better than the headline: 42% unrecovered becomes 0.7%, roughly sixty times smaller ^[15].

For that to transfer to your hardware, your failure mode has to look like a lost laser lock. It must be detectable from telemetry the agent already reads. The recovery must be a fixed, pre-approved procedure. A failed attempt must cost a retry rather than a sample. Relock clears all three. A BCA plate does not, because reagent and sample are consumed on the way to the plate reader ^[11]. So "real-time error handling" across a liquid handler and a robotic arm is the harder claim in the set, even though the laser is the one with a decimal point attached.

Janelia's weeks-to-a-day compression ^[13] arrives without a baseline protocol or instrument hours. If most of those weeks were queueing between vendor applications and operator handoffs, the result describes their scheduling rather than their microscopes, and it transfers only to labs with the same handoff structure. "Weeks" is a range.

Anthropic says safety checks and human approvals are retained for higher-risk decisions ^[4], which is the correct shape. The classification is the whole product. And the public announcement does not describe the technical controls, the permissions model, or the device-level safeguards participants use ^[10], nor does it establish a measured integration-time reduction for any device class ^[16]. A lab in the cohort therefore cannot assemble a rig-level safety case from public material. It negotiates one.

In my context I would take the UST-shaped surface first: interpret hardware documentation, generate and run tests, diff live readings against a digital twin to find regressions ^[14]. The output there is a comparison, and a wrong comparison costs a review rather than a plate. Actuation can wait for the permissions model to be written down.