Our Portfolio company QuEra. just automated one of the most time-consuming and boring parts of running a quantum computer using Anthropic’s new hardware standard.
When a quantum computer is delivered to a customer, the work has only just started.
The machine needs constant attention: lasers must be calibrated, optics adjusted, and frequency drift corrected. This also requires people with the right skills.
In a university lab, this knowledge is passed from one PhD student to another. If the laser loses its lock at 2 a.m., someone has to go to the lab and fix it. For one lab, this is an inconvenience, but manageable. For a company operating many machines at customer sites, it becomes a real limit to scaling.
Here is a specific example. A titanium-sapphire (Ti:Sa) laser needs to keep its frequency accurate to roughly one part in a trillion (super high level of precision!). Temperature changes, vibrations, or even opening a lab door can cause the frequency to drift. An engineer may need 5 to 10 minutes to fix it. By then, a quantum computation with error correction may already be ruined.
This is how the industry works today. But we may now be seeing the beginning of serious automation in this area. Last week Anthropic introduced MHS (Model Hardware Standard), a standard for connecting AI agents to hardware. It works like MCP, but for physical equipment: the agent can read, write, and see the hardware limits in a standard way.
QuEra built a driver for its laser and gave an agent a simple task: write a Python script to relock the laser frequency, with a success criterion of relocking on the first attempt and keeping the lock for 30 seconds.
The process is simple. The agent proposes a hypothesis, writes a normal script, and runs it on the real laser. The result is recorded in a log, which the agent reads before starting the next iteration. Importantly, there is no AI in the hardware control loop. The actual laser is controlled by deterministic code.
Over one night, the system ran about 760 tests. It started at 150 seconds per relock with a 58% success rate, using a script that had previously taken 4 people several months to develop. By 11 p.m., it was down to 22 seconds. By morning, it reached 6 seconds with a 96% success rate. In just one night, the success rate increased from 58% to 96%.
The final blind test was even better: 695 successful relocks out of 700, or 99.3%.This is a very impressive result. Of course, a quantum computer contains a huge amount of equipment, and the Ti:Sa laser is just one of hundreds of components. But this is a strong starting point. We believe the bottleneck can start to shift. Instead of experts manually adjusting the hardware, the expert defines a task, while a script handles the adjustments. A script that can be reviewed, understood, and deployed without AI in the control loop.
