Four people spent several months building Quer’s laser recovery script.
Success rate: 58%
Recovery time: about 150 seconds
Then four Claude instances worked on the live laser in a loop. By morning:
Success rate: 96%
Recovery time: 6 seconds
The stronger result came after Quer removed the agents.
The completed Python script delivered 695 clean recoveries in 700 randomized trials. It was deterministic, inspectable, and ran without a model.
The agents helped build the automation. They did not have to remain in control.
Two Actions Connected Incompatible Machines
Anthropic published its model hardware standard on August 27 as a research preview, after four labs tested it on real hardware.
Every device exposes two actions: Read and Write. It also describes its measurements, capabilities, and safety limits in a standard format that the model reads before operating it.
The interface can work through a tool protocol, command line, or code files. For fast operations, an agent can write a command sequence that runs without model reasoning between every step.
It is model agnostic and does not require Claude.
At Carnegie Melon, this connected equipment controlled through job files, an old window scripting interface, and a screen-and-mouse interface.
The Agents Could Recover Experiments
At Genentech, an agent coordinated a liquid handler, robot arm, and plate reader for a protein assay.
Its first attempt treated every liquid the same. The thick protein foamed.
The agent tested dyed-liquid transfers, measured the results, compared them with a human expert’s transfer, and changed the settings:
- Water: 140 microL a second
- Thick protein: 10 microL a second
Genentech’s automation experts said both settings were sensible.
At Carnegie Melon, researchers deliberately introduced six faults, including a missing plate, a rotated plate, an unplugged camera, and an emergency stop. All six were blocked before any equipment moved.
The agent also rejected a poor dose response curve, discarded the plate, and repeated the experiment with half the top concentration. The complete eight-hour workflow, including the second attempt, was about three times faster.
In another case, a camera found foam in a tube held by a robot arm. The software found a connected centrifuge, proposed using it through Slack, and spun the tube to remove the bubbles.
These systems were responding to faults and experimental results, not merely running fixed equipment sequences.
Agent Search Became Inspectable Automation
Quer separated exploration from operation.
Four Claude instances proposed changes, wrote code, tested the live laser, recorded the results, and chose the next attempt.
The finished script did not blindly replay a fixed sequence. It inspected the instruments and changed only the controls that were wrong.
Across 363 experiments and 16 unattended hours, residual error fell from 15.7 molts to 1.55.
A specialist independently retuned the same laser. The settings were similar across almost the entire tested band, although the human tune had about a thousand times more noise at one resonance.
During a 19-hour run, the agent’s settings never lost the lock. The experts lost it about one and a half times an hour.
This does not make the agent a universal replacement for expert control. It shows that agent-guided experimentation can produce ordinary, inspectable automation without a live model dependency.
Six Instruments, Robot Arms, and a Room-Sized Microscope
At the University of Washington, Zi Song connected six instruments in under a week, including driver development.
An agent watched DNA copying curves, asked whether it should stop, then halted the run and lowered the machine to 4° when instructed.
Song also connected an open source robot arm built on Hugging Face’s lay robot to a liquid handler. The arm moved 10 seconds after dispensing, and repeated runs avoided collisions.
At Janellia, every instrument on Virginel Ruten’s room-sized microscope rig writes into a shared dictionary. An agent can inspect images as they form, find an interesting region, and zoom in during the experiment.
Adoption Signs Do Not Make It a Standard
Eight hardware companies are building support into their instruments. The two biggest names in open robotics are adding it to their libraries.
Those are signs of adoption, but they do not settle whether the interface is a standard.
Access to the research preview requires an application. Anthropic says an open source release will follow safety evaluations, and the speaker cautioned that the word “standard” has not earned its role yet.
The Physical World Still Pushes Back
The Genentech agent could adjust from experimental feedback, but its physical understanding was weak.
When it saw bubbles, it read the error code and treated them as a software fault. Its view of the physical world came from text.
Anthropic says spatial and physical reasoning still needs expert oversight.
Safety controls also slowed the work. Some experiments stopped overnight because the agent waited for approval before taking actions it considered slightly risky.
That delay has to be considered alongside what these systems control: robot arms, centrifuges, and experimental settings.
Safety Files May Become Regulated Components
On January 20, 2027, a new European machinery regulation replaces the previous directive.
It covers AI-based safety functions and machines that can change their own behavior. A file containing a robot arm’s safety limits could therefore become a regulated safety component.
More Agents Can Also Mean More Waiting
A criticism from @Cloud-Codes focused on orchestration overhead.
In its example, a language model named six calls to other models only three lines into a program. Each request took seconds while the harness watched and nothing happened.
More model calls can mean more waiting. They do not, by themselves, show that useful work is moving forward.
Quer is the stronger counterexample because its loop produced a concrete result that became deterministic code.
The Hiring Note Did Not Provide Enough Detail
A Tamil post from @aicoachjohn advised people seeking AI work to use the resources available to them.
It mentioned a 2018 computer science engineering graduate who did well and was selected. The excerpt did not clearly explain the three rules promised in its title, so there is not enough detail to take the point further.
The lab results leave a harder question: When agent-written code controls regulated machinery, who is responsible for a failure if no person wrote or reviewed it?