Claude-developed controller fixes quantum computer laser drift in seconds

A quantum computer can have hundreds of qubits, but one drifting laser can still bring the whole machine to a halt. Keeping those lasers precisely tuned is one of the less glamorous, and surprisingly difficult parts of running a quantum computer. Now, researchers have shown that an AI agent can learn to diagnose and correct some of these failures on its own. In a test involving a neutral-atom quantum-computing system operated by Massachusetts-based QuEra Computing, Anthropic’s Claude restored a laser’s lock in 695 of 700 trials, usually in less than six seconds. The result points to a different role for AI in quantum computing: not solving a quantum problem, but learning how to keep the hardware itself working. “For years, the hardest part of scaling quantum computers wasn’t the physics. It was the people driving at 2 am to fix a laser lock. We built a solution using the Model Hardware Standard to fix that,” Sergio Cantu, VP of Quantum Systems at QuEra Computing, said. Teaching an AI to troubleshoot a quantum machine Neutral-atom quantum computers use lasers to manipulate individual atoms and use them as qubits. The lasers need to remain at extremely precise frequencies. Disturbances can make them drift, causing them to lose their lock and interrupt the computer’s operation. This precision is also central to neutral-atom quantum operations, where laser light is used to trap and manipulate atoms. Some common laser problems can already be corrected automatically. The difficult cases are those that require an expert to examine measurements, figure out what went wrong, and decide which setting to adjust. There is no simple troubleshooting recipe that covers every possible failure. Previously, four specialists spent two to three weeks creating a recovery script by hand. But that approach had an obvious limitation: the script could only handle failures its authors had anticipated. The researchers instead gave Claude access to a dedicated laser testbed through the Model Hardware Standard, or MHS. Developed by Anthropic and HHMI Janelia Research Campus, MHS is designed to let AI agents interact with scientific equipment while keeping operations within predefined safety limits. Rather than handing Claude a troubleshooting checklist, the researchers allowed it to experiment. They introduced different disturbances, after which the AI could change a setting, measure the result and decide what to try next. It repeated this cycle across hundreds of failure conditions, including overnight, gradually finding combinations of settings that could restore the laser’s lock. The final result was conventional control software rather than an AI model making decisions every time the laser failed. Engineers reviewed the process and defined what constituted success, while hardware safeguards such as operating limits, interlocks and emergency stops restricted what the agent could do. This distinction is important. Claude helped discover the recovery strategy, but the resulting controller could then run deterministically without an AI model making decisions inside the control loop. From overnight experiments to six-second fixes The resulting controller was tested against seven types of faults. It restored the laser successfully in 695 of 700 timed trials. It also never reported success when the laser had not actually recovered. The five unsuccessful trials shared a condition of the test rig rather than indicating a failure of the recovery logic. The speed of recovery was another major difference. Most faults were cleared in less than six seconds, while the hardest cases took roughly 10 to 14 seconds. A human specialist typically needed about five to 10 minutes for comparable recovery work. Claude was also asked to improve the laser’s performance rather than simply recover it. Its tuning reduced residual noise by a factor of five and prevented the system from dropping out during unattended operation. When the settings were later checked with an independent instrument that the AI could not influence, they matched those produced by an experienced specialist and corrected a flaw that the manual tuning had missed. The researchers then tested whether the approach could transfer to another laser wavelength. Claude worked out the required settings from scratch during one unattended overnight run—a process that would normally take weeks of hands-on commissioning. This kind of automation could become increasingly useful as quantum hardware moves toward larger-scale, deployable systems, where lasers, control electronics and other precision components all have to work reliably together. A step toward self-maintaining quantum hardware The experiment does not mean quantum computers can now repair themselves. It was conducted on a dedicated testbed, and human engineers still established the safety boundaries and judged the results. The system also cannot solve physical hardware failures simply by changing software settings. The pilot covered one subsystem, rather than demonstrating autonomous maintenance of an entire quantum computer. However, the experiment addresses a growing problem as quantum computers become larger and more widely deployed. Every additional precision subsystem creates more opportunities for faults, while customer machines may be located far from specialists who know how to fix them. “A customer expects the entire computer, and thus every subsystem, to hold itself together without a specialist in the room. This is why the results from the MHS research preview and Anthropic’s frontier AI models are so meaningful. We are making it far easier and cheaper to keep our computers running at their best,” Takuya Kitagawa, President of QuEra, said. Similar efforts to make quantum hardware more resilient include quantum chips designed to verify their own hardware, showing how reliability is becoming a problem that researchers are tackling at several levels. The broader test is whether AI agents can learn similar troubleshooting and calibration tasks for other scientific hardware. Researchers now need to establish whether the approach remains reliable on operating quantum computers and across different machines. If it does, AI could become less a tool for performing calculations and more a layer of automation that helps keep complex scientific instruments running. Get the latest in engineering, tech, space & science - delivered daily to your inbox.Rupendra Brahambhatt is an experienced writer, researcher, journalist, and filmmaker. With a B.Sc (Hons.) in Science and PGJMC in Mass Communications, he has been actively working with some of the most innovative brands, news agencies, digital magazines, documentary filmmakers, and nonprofits from different parts of the globe. As an author, he works with a vision to bring forward the right information and encourage a constructive mindset among the masses.