Anthropic's Claude agent writes self-repairing laser software for QuEra quantum computers

AI agent takes over quantum laser maintenance

QuEra Computing, a Boston-based maker of neutral-atom quantum computers, used an AI agent built around Anthropic's Claude model to produce software that automatically repairs the laser systems at the heart of its machines. According to a report on dev.to, most failures handled by the new controller are resolved in under six seconds, where a human specialist would need up to ten minutes. The project is part of a broader push toward quantum hardware that can run at customer sites without a resident expert.

The problem: lasers that drift

Neutral-atom machines use precisely tuned lasers to control the atoms that act as qubits. Small environmental changes — passing foot traffic, temperature swings — can push laser frequencies off target and halt calculations until someone intervenes. Per dev.to, routine corrections were already automated, but complicated failures still had to be escalated to the people who originally designed the system.

How the fix was built

The work relied on the Model Hardware Standard, a framework jointly developed by Anthropic and the Janelia Research Campus that lets AI models operate physical lab equipment within hard safety limits, including emergency-stop rules. Claude worked through hundreds of failure scenarios on a testbed and produced a permanent control program.

Notably, the deliverable is conventional software rather than an AI left permanently in charge: engineers can read, audit and verify the generated code, so the recovery logic stays transparent during live operation.

Test results

In testing reported by dev.to, the controller ran 700 trials across seven fault categories and restored the laser 695 times. The remaining failures were attributed to the test equipment itself rather than the controller's logic. Most recoveries finished in fewer than six seconds, and the system never falsely claimed a successful repair — a property dev.to calls essential for reliability. The testbed was not isolated: it sat in an active facility with people and temperature changes nearby, and the controller coped unaided throughout the pilot.

The agent also went beyond repair. It found settings that cut background noise by 80 percent compared with earlier methods, matching the work of veteran physicists and catching subtle errors humans had overlooked. An independent measuring instrument confirmed the machine-generated settings beat the manual baseline. Later, tasked with configuring a second laser system, the agent finished the job in an unattended overnight run — work that would normally take an engineer weeks of hands-on effort.

A scaling problem, not a physics one

As quantum computers move from research labs into corporate data centers, maintenance becomes a logistics bottleneck. Larger machines carry more lasers and therefore more chances of a fault, and when hardware sits far from the manufacturer, dev.to notes that manual tuning can take around 30 minutes and may require travel or off-hours work.

QuEra intends to extend the approach across other subsystems, with a stated goal of machines needing no manual intervention for basic operations. That ambition ties directly into its roadmap: the company is coordinating with Amazon Web Services to offer cloud access to its upcoming Libra system by 2028, and its partnerships with Hewlett Packard Enterprise and NVIDIA assume quantum systems can behave like ordinary data-center hardware.

The Model Hardware Standard itself remains in a research phase, focused on keeping AI control over physical equipment safe, and other organizations can apply for access.

Why it matters

The notable part is the division of labor. An AI agent explored a physical system, generated auditable control code, and outperformed expert manual tuning — while humans stay out of the loop at runtime but keep full visibility into the logic. If the pattern generalizes, hardware vendors could commission and maintain complex equipment by directing agents at it rather than scripting every failure mode by hand; dev.to reports that bringing up new laser wavelengths has dropped from weeks to a single night. For quantum computing specifically, subsystems that recover on their own look like a precondition for the enterprise deployments vendors keep promising, and this is one of the clearer demonstrations so far of a frontier AI model doing safety-critical engineering on real lab hardware.