On August 27, developer of quantum computers QuEra Computing said that Anthropic’s AI agent Claude developed and tested on a separate test stand a program to automatically restore a laser’s operating frequency. In a control series of 700 tests across seven types of failures, the controller it created successfully returned the setup to the target state 695 times.
The system never reported a successful recovery when one had not actually occurred. QuEra attributed the five unsuccessful attempts to the state of the experimental setup rather than a software error.
QuEra’s quantum computers use neutral atoms as qubits. Nearly all operations to control atomic qubits and read out their state are performed via interactions of the atoms with laser light.
Temperature, vibrations, and pressure changes can break the laser’s frequency lock. In such cases, quantum operations begin to fail, and restoring more complex faults requires a specialist’s intervention.
According to QuEra, an experienced operator typically needs five to ten minutes to return the laser to its operating frequency. The company had previously automated simple disruptions, but rare and more complex scenarios still required manual intervention.
Claude rewrote the recovery logic
For the experiment, QuEra used the Model Hardware Standard (MHS), a specification for interfacing AI agents with physical equipment. Its development began as a joint project by Anthropic and the HHMI Janelia research center.
MHS gives an agent standardized access to equipment readings and controls while preserving engineer-defined constraints, hardware interlocks, and an emergency stop.
The technology is currently in a limited research preview, but Anthropic plans to open-source the standard. QuEra connected Claude through MHS to a separate test stand with precision equipment costing about $700,000.
The workflow was split among four roles, each performed by a new instance of Claude. One proposed a hypothesis, the second made code changes, the third ran the program on the equipment and recorded results, and the fourth analyzed the log and chose the next change. Engineers defined the experiment’s boundaries and success criteria and reviewed each stage.
According to Anthropic, the cycle repeated hundreds of times overnight.
A script previously created by the QuEra team restored the laser in about 58% of cases and took roughly 150 seconds per attempt. During development, Claude raised the rate to 96% and cut the time to about six seconds.
After development, the program was tested separately without the AI agent. In 695 of 700 trials it returned the laser to the required frequency, a 99.3% success rate.
For failures without a hop to another wavelength, recovery took from 0.9 to 5.4 seconds. The most complex cases required about 10–14 seconds, versus five to ten minutes for a specialist.
In the lab, the controller handled 43 failures
QuEra placed the stand in a working laboratory, so the equipment continued to experience people moving around and other external disturbances. During the pilot, the laser spontaneously lost the required mode 43 times. In all cases, the controller restored it without human intervention.
At the same time, in normal operation recovery is not handled by Claude. AI was used to develop and test the algorithm, and the stand runs a standard deterministic program with fully auditable code.
“It is important to draw a distinction: MHS is the environment in which Claude designed, wrote, and verified the controller. The stand runs a standard deterministic program with fully auditable code,” — emphasized QuEra.
The next step was not only recovery after failures but also improving the system’s stability itself. The quality of the frequency lock depends on 12 interrelated feedback parameters. Claude was allowed to change them, measure the resulting noise, and search for an optimal combination.
Anthropic’s technical breakdown says that over 16 hours the agent conducted 363 experiments. The RMS residual error fell from 15.7 to 1.55 mV.
With parameters chosen by Claude, the system did not lose the lock once over 19 hours. With the set of parameters manually selected by a specialist, this occurred on average about 1.6 times per hour.
For additional verification, Claude’s and the specialist’s parameter sets were compared on a separate phase-noise analyzer, to which the agent did not have access. Overall, the results were comparable to manual tuning by an experienced specialist. Claude’s configuration additionally suppressed by roughly 1,000 times the resonant noise around 220 kHz that remained after manual tuning.
Unlike automatic recovery, at the tuning stage Claude remains directly involved in the experimental process. QuEra plans to package the procedure into a separate tool.
The approach was also tested on a laser with a different operating wavelength. The agent selected parameters in a single autonomous overnight run. The company claims that similar manual preparation usually takes weeks.
Limitations
The pilot covered one laser system on a dedicated stand. Porting the controller to QuEra’s production quantum processors is only planned. Other components of the setup still require calibration, control, and repair.
Claude’s work required a significant amount of initial context and constant oversight by engineers. QuEra noted that a specialist stopped the agent several times when it chose a seemingly plausible but incorrect path.
Anthropic also noted that Claude struggled when the problem arose in the physical hardware, since its understanding of the setup was based on software data. For potentially risky actions, the model could also pause and request human confirmation.
Next, QuEra intends to port the recovery controller to production quantum processors, create a separate tuning tool, and test the approach on other subsystems. The company views maintenance automation as one way to reduce future commercial systems’ dependence on on-site specialists.
In July, in its new “Quantum & After” column, ForkLog explained whether it is possible to make money on quantum technologies.
