On August 27, QuEra Computing, a company specializing in the development of quantum computers, presented the results of a landmark experiment: Anthropic's AI agent Claude independently developed and tested a program for automatically restoring the operating frequency of a laser. During control tests, which included 700 tests across seven types of failures, the created agent successfully returned the setup to its target state in 695 cases, amounting to an impressive 99.3% efficiency.
Notably, the system never gave a false signal of successful recovery. QuEra attributes the five failed attempts to the state of the experimental setup, not errors in the program code.
The Critical Role of Lasers in Quantum Computing
QuEra's quantum computers use neutral atoms as qubits, and virtually all operations for controlling these qubits, as well as reading their state, are carried out through laser radiation. Temperature fluctuations, vibrations, and pressure changes can disrupt the laser's frequency lock, leading to failures in quantum operations. Recovering from complex failures typically requires the intervention of an experienced specialist, who needs five to ten minutes.
How Claude Created the Controller
For this experiment, QuEra used the Model Hardware Standard (MHS) specification, developed jointly by Anthropic and the HHMI Janelia research center. MHS provides the agent with standardized access to hardware while maintaining engineering constraints and emergency shutdown. Claude was connected to a separate test stand worth approximately $700,000.
The workflow was divided among four roles, each performed by a separate instance of Claude: one generated hypotheses, a second made code changes, a third ran the program and recorded results, and a fourth analyzed logs. Engineers monitored each stage. A pre-existing QuEra script recovered the laser in only 58% of cases, spending about 150 seconds per attempt. Claude, however, improved this figure to 96%, reducing the time to six seconds.
Superiority Over Manual Control
After development, the program was tested without AI involvement: in 695 out of 700 trials, it returned the laser to the correct frequency. For simple failures, recovery took from 0.9 to 5.4 seconds, and in complex cases, 10-14 seconds, whereas a specialist required five to ten minutes.
In a working laboratory environment, where the equipment was subject to external disturbances, the controller successfully handled 43 spontaneous laser failures. It is important to emphasize that the test stand runs an ordinary deterministic program, and Claude was used only for its development and testing.
Parameter Optimization and Future Prospects
In the next phase, Claude worked on improving system stability by modifying 12 interconnected feedback parameters. Over 16 hours, the agent conducted 363 experiments, reducing the RMS residual error from 15.7 to 1.55 mV. With the parameters selected by Claude, the system did not lose lock once over 19 hours, whereas with a specialist's manual tuning, this occurred on average 1.6 times per hour. Moreover, Claude's configuration suppressed resonant noise around 220 kHz approximately 1000 times better than manual tuning.
The experiment was also successfully repeated on a laser with a different operating wavelength: the agent selected parameters overnight, whereas manual preparation typically takes weeks.
Limitations and My Analysis
Despite the impressive results, the pilot was limited to a single laser system, and transferring the controller to operational quantum processors is only planned. Claude's work required significant oversight from engineers, who stopped the agent several times when it chose a wrong direction. Additionally, the AI struggled with issues arising directly in the physical equipment, as its representation was based on software data.
Nevertheless, this experiment demonstrates the enormous potential of AI in automating complex physical systems. The agent's ability not only to speed up the process by dozens of times but also to surpass humans in tuning accuracy opens new horizons for the commercialization of quantum technologies. QuEra plans to transfer the controller to operational processors and test the approach on other subsystems, which could significantly reduce dependence on narrow specialists in the future.