Quantum computing developer QuEra Computing has reported the successful application of neural network technologies to automate complex physical processes. Claude, an intelligent agent created by Anthropic, independently designed and tested on a dedicated bench a program capable of automatically returning a laser system to its target operating frequency. During a control series of seven hundred trials covering seven different types of hardware failures, the generated controller successfully restored the equipment to working order in 695 cases. The system recorded zero false positives, and the five failed attempts were attributed by engineers to the physical condition of the experimental bench rather than logical errors in the written code.
The architecture of QuEra's quantum computers is based on the use of neutral atoms as qubits, with virtually all operations for controlling and reading their state carried out via precise laser radiation. External factors such as temperature fluctuations, vibrations, and changes in atmospheric pressure can disrupt laser frequency locking, which inevitably leads to failures in quantum computations. While the company had previously automated simple deviations, remedying complex failures has traditionally required manual intervention. An experienced operator engineer typically needs five to ten minutes to return the system to a stable state, creating a bottleneck in the maintenance of costly equipment.
To integrate the neural network with the physical world, the Model Hardware Standard (MHS) was used — a protocol developed jointly by specialists from Anthropic and the HHMI Janelia research center. This protocol provides intelligent agents with unified access to sensors and hardware controls while maintaining strict engineering constraints, hardware interlocks, and emergency stop functions. As part of the experiment, QuEra connected Claude via MHS to an isolated bench housing precision equipment valued at approximately $700,000. The working cycle was distributed among four independent instances of the neural network: the first proposed hypotheses, the second modified the program code, the third executed the program and logged results, and the fourth analyzed the data and planned the next step. Engineers set the boundaries of the experiment and verified each stage, while the optimization cycle itself was repeated hundreds of times over the course of a single night.
The effectiveness of the solution proposed by the neural network far surpassed previous achievements. A script previously created by the QuEra team managed to recover the laser in only 58 percent of cases, taking approximately 150 seconds per attempt. The algorithm generated by Claude raised the success rate to 96 percent, reducing response time to approximately six seconds. For failures without a jump to a different wavelength, recovery took between 0.9 and 5.4 seconds, while the most complex scenarios required 10–14 seconds. After the development phase was complete, the program was tested in autonomous mode without neural network involvement. Placing the bench in an active laboratory where staff movement created additional disturbances, researchers recorded 43 spontaneous losses of laser lock. In all cases, the deterministic controller written by the neural network returned the system to normal without any human intervention. QuEra representatives specifically emphasized that what runs on the actual hardware is not the neural network itself, but ordinary verifiable program code created by the intelligent agent within the MHS environment.
The next stage of the research involved optimizing system stability by tuning twelve interdependent feedback parameters. Over sixteen hours, the agent conducted 363 independent experiments, measuring noise levels and searching for the ideal configuration. As a result, the root mean square (RMS) residual error decreased from 15.7 to 1.55 millivolts. With the parameters selected by the neural network, the system did not lose frequency lock once during nineteen hours of continuous operation, whereas with manual tuning by a specialist, similar dropouts occurred on average 1.6 times per hour. Additional verification on an isolated phase noise analyzer, to which the agent had no access, confirmed that the results were comparable to the work of an experienced engineer. Furthermore, the Claude configuration additionally suppressed resonance noise around 220 kilohertz approximately a thousand times more effectively than manual calibration. A similar approach was successfully tested on a laser with a different operating wavelength: an autonomous overnight run allowed parameters to be selected in the time that specialists typically spend weeks on.
Despite the impressive results, the current pilot project covered only one laser system on a dedicated test bench, and transferring the controller to QuEra's active quantum processors is still only planned. The remaining components of the complex systems still require manual calibration and repair. The code generation process required providing the neural network with extensive context and constant supervision: engineers had to stop the agent several times when it chose a path that was superficially logical but physically incorrect. In addition, the model experienced difficulties diagnosing hardware failures based solely on software telemetry data, and paused to request human confirmation before potentially risky actions. Going forward, QuEra intends to integrate the developed algorithms into commercial quantum processors and create dedicated tools for automatic tuning, which will reduce the dependence of technology companies on scarce, highly specialized on-site experts. Industry publications have previously also noted growing interest in the commercial potential of such quantum technologies.