A Laser Recovery Rate Went From 58% to 99.3%. That Is the Story.

On 27 August, Anthropic released the Model Hardware Standard as a research preview: a specification for AI agents to discover, operate, and troubleshoot physical equipment. Microscopes, liquid handlers, robotic arms. Partners at launch included AWS, Universal Robots, and Hugging Face, with open-sourcing planned after further safety evaluation.

I wrote last week about the agent plumbing settling as standards consolidated. This is the same pattern one layer down, and ordinarily I would not write about a second protocol in two weeks. The pilot results are the reason to.

The numbers from the pilots

A drug discovery assay at Genentech. An imaging workflow at Janelia that went from weeks to a day. And laser recovery at QuEra, a quantum computing company, improving from a 58% success rate to 99.3%.

That last figure is the one to sit with. It is not a productivity gain. Going from 58% to 99.3% is the difference between a process that needs a person watching it and a process that does not, and those are different categories of thing rather than different points on a scale.

What was actually broken

“Going from 58% to 99.3% is the difference between a process that needs a person watching it and a process that does not. Those are different categories of thing, not different points on a scale.”

Notice what the constraint was not. It was not that nobody at QuEra understood laser calibration. The expertise was present, and the equipment worked. The gap was between knowing what to do and having something available to do it consistently, at volume, at three in the morning, without getting tired.

That gap is everywhere in physical industries and it has been invisible because there was no realistic alternative. Every laboratory, plant, and workshop has a set of procedures that are executed by people at a quality level that varies with attention, staffing, and how long the shift has been.

Making that layer addressable does not require any new knowledge about the underlying process. It requires the process to be written down precisely enough for something else to run it, which is a different and much more boring problem.

Going from 58% to 99.3% is the difference between a process that needs a person watching it and a process that does not.

Why this favours people who have run the equipment

Writing a procedure down precisely enough is where this becomes about experience rather than technology.

The reason plant procedures are not already specified to that standard is that humans do not need them to be. A person who has run the machine for six years fills the gaps automatically: they know that the reading drifts when the room is cold, that a particular noise means stop, that step four is nominally optional and never actually is. None of that is in the manual. It is in them.

An agent operating the same equipment needs all of it made explicit, and the only people who can supply it are the ones carrying it. That is a genuinely scarce input, it is concentrated in industries where nobody is currently building software, and it does not transfer between them.

The honest timeline

This is a research preview with a handful of pilots, not a product, and the industries it points at move slowly for reasons that are usually good. Nothing here suggests anyone should start building against it this quarter.

What it does suggest is where to point your attention if your background is operational rather than digital. The software layer over physical work has just acquired a plausible standard and serious backing, and the people positioned for what comes next are the ones who can specify a physical process precisely. If that describes your last decade, this is the most interesting thing that happened this month.

See how to find a vertical in an ignored industry for choosing where to look, and industry knowledge is the moat for why the specification is the asset. The system covers structuring the search.