On August 27, QuEra Computing announced that its AI agent, Claude from Anthropic, has developed and tested a program designed to automatically restore the operational frequency of lasers. In a controlled series of 700 tests involving seven types of failures, the controller created by Claude successfully returned the setup to its target state 695 times.

The system never reported a successful recovery when it did not actually occur. QuEra attributed the five unsuccessful attempts to the condition of the experimental setup rather than a flaw in the program.

QuEra's quantum computers utilize neutral atoms as qubits, and nearly all operations for managing these atomic qubits and reading their states are conducted through interactions with laser light.

Factors like temperature, vibrations, and pressure changes can disrupt the laser frequency fixation. When this happens, quantum operations may fail, and restoring more complex failures typically requires a specialist's intervention.

According to QuEra, an experienced operator usually needs five to ten minutes to return the laser to its operational frequency. While the company has previously automated simpler disruptions, rare and more complex scenarios still required manual intervention.

Claude Redefined Restoration Logic

For the experiment, QuEra utilized the Model Hardware Standard (MHS), a specification for AI agents interacting with physical equipment. This standard was developed as a collaborative project between Anthropic and the HHMI Janelia research center.

MHS provides standardized access for the agent to equipment readings and controls while maintaining constraints set by engineers, including hardware locks and emergency stops.

The technology is currently in a limited research preview, but Anthropic plans to eventually release the source code for the standard. QuEra connected Claude to a separate setup with precision equipment valued at approximately $700,000 via MHS.

The workflow was divided among four roles, each handled by a new instance of Claude. One instance proposed hypotheses, another modified the code, a third executed the program on the equipment and recorded results, while the fourth analyzed logs and determined the next change. Engineers defined the experiment's boundaries and success criteria, verifying each stage.

Diagram of Claude's interaction with the laser system via MHS. Source: Anthropic.

As reported by Anthropic, the cycle was repeated hundreds of times over the course of the night.

A script previously created by the QuEra team restored the laser in about 58% of cases and took approximately 150 seconds per attempt. Through Claude's development, this rate increased to 96%, while the time per attempt decreased to around six seconds.

Source: Anthropic.

After development, the program was tested independently without AI agent involvement. It successfully returned the laser to the required frequency in 695 out of 700 trials, achieving an effectiveness of 99.3%.

In cases of failures without a wavelength shift, restoration took between 0.9 to 5.4 seconds. The most complex cases required about 10 to 14 seconds compared to the five to ten minutes typically needed by a specialist.

Controller Managed 43 Failures in the Lab

QuEra installed the setup in an active laboratory, so the equipment continued to be affected by people moving around and other external disturbances. During the pilot, the laser spontaneously lost its required mode 43 times, but the controller restored it in every instance without human intervention.

However, in operational mode, Claude does not perform the restoration. The AI was used for developing and testing the algorithm, while a standard deterministic program with fully verifiable code operates on the setup.

"It's important to distinguish: MHS is the environment in which Claude designed, wrote, and tested the controller. The deterministic program with fully verifiable code operates on the setup," emphasized QuEra.

The next step involves not only recovery from failures but also enhancing the stability of the system itself. The quality of frequency fixation depends on 12 interrelated feedback parameters. Claude was granted the ability to modify these parameters, measure the resulting noise, and seek optimal combinations.

According to a technical breakdown by Anthropic, over 16 hours, the agent conducted 363 experiments. The RMS residual error decreased from 15.7 to 1.55 mV.

With the parameters selected by Claude, the system did not lose fixation once over a span of 19 hours. In contrast, with parameters manually selected by a specialist, this occurred on average about 1.6 times per hour.

To further validate, sets of parameters from Claude and the specialist were compared on a separate phase noise analyzer, which the agent did not have access to. Overall, the results were comparable to the manual tuning done by an experienced specialist. Claude's configuration further suppressed resonant noise around 220 kHz, which remained after manual adjustment, by about 1000 times.

Optimization of laser parameters by Claude: RMS error reduction from 15.7 to 1.55 mV. Source: Anthropic.

Unlike the automatic restoration, Claude remains directly involved in the experimental process during the tuning stage. QuEra plans to package this procedure into a separate tool.

This approach was also tested on a laser with a different operational wavelength. The agent adjusted the parameters in a single autonomous overnight run. The company claims that similar manual preparation usually takes weeks.

Limitations

The pilot covered a single laser system on a dedicated setup. The transfer of the controller to operational quantum processors at QuEra is still planned. Other components of the setup still require calibration, monitoring, and repair.

Claude's work required a significant amount of initial context and constant oversight from engineers. QuEra noted that specialists had to stop the agent several times when it chose a seemingly plausible but incorrect direction.

Anthropic also noted that Claude faced challenges when problems arose directly from the physical equipment, as its understanding of the setup was based on software data. In potentially risky scenarios, the model could halt and request human confirmation.

QuEra's next steps include transferring the recovery controller to operational quantum processors, creating a separate tuning tool, and testing the approach on other subsystems. The company views maintenance automation as a way to reduce future commercial installations' dependence on niche specialists on-site.

As a reminder, in July, ForkLog explained in the new section "Quantum & After" whether it's possible to profit from quantum technologies.