Quantum computing stands on the brink of a new era, and it's not just about qubit power, but also about their maintenance. QuEra Computing, a leader in developing quantum systems based on neutral atoms, has presented the results of an experiment in which Anthropic's AI agent Claude didn't just assist, but independently created and tested a program for automatically restoring the operating frequency of a laser. This is an important step toward reducing the dependence of commercial quantum setups on the manual labor of highly skilled engineers.
The essence of the problem is simple: in QuEra's quantum computers, lasers control atom-qubits, and any frequency deviation—due to temperature, vibrations, or pressure—leads to failures. Previously, recovering from complex failures took an experienced operator five to ten minutes. Now, during a control series of 700 trials, the controller created by Claude successfully returned the laser to its target state 695 times, demonstrating a 99.3% success rate.
How AI Rewrote the Rules of the Game
For the experiment, QuEra used the Model Hardware Standard (MHS) specification, developed jointly with Anthropic and the HHMI Janelia research center. MHS gives the agent standardized but safe access to equipment, preserving hardware locks and emergency stop. The workflow was divided into four roles performed by separate instances of Claude: one generated hypotheses, another wrote code, a third ran the program, and a fourth analyzed the results. Engineers only set the boundaries of the experiment and monitored each stage.
The result is impressive. QuEra's previously existing script restored the laser in 58% of cases, spending about 150 seconds per attempt. Claude improved this figure to 96%, reducing the time to six seconds. In complex cases, recovery took 10–14 seconds, which is still several times faster than manual work by a specialist.
Calibration as an Art
But the experiment didn't end there. QuEra set a more ambitious task for the AI—not just to fix failures, but to improve the stability of the entire system. The quality of frequency locking depends on 12 interrelated feedback parameters. Over 16 hours, the agent conducted 363 experiments, reducing the RMS residual error from 15.7 to 1.55 mV. With the parameters selected by Claude, the system didn't lose lock once over 19 hours, whereas with manual tuning, this happened on average 1.6 times per hour.
Notably, Claude didn't just copy the engineer's actions. It additionally suppressed resonant noise around 220 kHz by approximately 1000 times, which went unnoticed during manual tuning. This suggests that AI is capable of finding non-obvious solutions that might elude a human.
Limitations and Prospects
It's worth noting that the pilot was conducted on a separate test bench, not on operational quantum processors. The agent's work required constant supervision, and engineers stopped it several times when it chose a seemingly plausible but erroneous direction. Additionally, Claude struggled with issues arising directly in physical equipment, as its understanding is based on software data. Nevertheless, the next stage is transferring the controller to real systems and creating a separate tuning tool.
My view: This experiment is not just another demonstration of AI capabilities. It's a signal that the automation of complex physical systems is becoming a reality. AI's ability not only to restore but also to optimize parameters faster and more accurately than a human opens the path to larger-scale and more reliable quantum setups. The question is how quickly we can trust AI with full control over such critical systems, and what new risks this will create.