QuEra Computing, a developer of neutral-atom quantum computers, announced on August 27, 2026, that it used Anthropic's Claude to develop laser recovery and tuning logic. The resulting recovery controller returned lasers to their target frequency in 695 out of 700 trials, a 99.3% success rate. But the notable shift here isn't that AI now continuously operates the quantum computer. It's that the knowledge Claude gained through trial and error on real hardware has been sealed into ordinary, inspectable code.
Lasers serve as the hands and feet of a neutral-atom machine. In QuEra's system, optical tweezers capture atoms, a separate laser manipulates their quantum states, and light readout ultimately yields computation results. According to Anthropic, the lasers involved here require frequency precision on the order of one part in a trillion. Temperature shifts, vibration, or air pressure changes can knock the system out of "lock," preventing correct operations from reaching the qubits. The task at hand was to compress a rare failure recovery process that human experts once handled in five to ten minutes down to a matter of seconds.
Claude builds it, but ordinary code runs it
What QuEra handed to Claude wasn't the entire customer-facing quantum computer, but a dedicated laser testbed worth roughly $700,000. The AI didn't operate inside the microsecond-scale servo loop itself; instead, it worked from a level above, reading equipment states and adjusting control values. Humans predefined the operating range and success criteria: the relock had to succeed on the first attempt and hold for 30 seconds.
The development loop was split into four roles: one Claude instance proposed new recovery hypotheses, another fixed the code, a third ran tests on real hardware and logged every operation, and a fourth decided the next change based on the logs. Each used a fresh instance, and by introducing beam interruptions, equipment power cuts, and frequency drift, the system ran through hundreds of improvement cycles overnight.
It's important to distinguish development from operation here. The recovery script doesn't call the AI at runtime, but the tuning workflow still keeps Claude in the active loop. The former is a standalone, deterministic program that Claude designed and validated—one whose logic humans can inspect. It's not a setup where the system waits for a large language model's response during a failure, or where the model makes a different judgment call each time.
Breaking down the 700 trials into three stages
The path to 99.3% involves three distinct sets of numbers. According to Anthropic, the conventional fixed script—which simply replayed the human operator's procedure step by step—achieved roughly a 58% recovery rate and took about 150 seconds. Its weakness: if room temperature or air pressure shifted mid-procedure, steps already completed became invalid, forcing a restart from the beginning.
| Stage | Reported success rate | Recovery time | What it measures |
|---|---|---|---|
| Conventional fixed script | ~58% | ~150 seconds | Automating a human's linear procedure |
| Unsupervised development run by Claude | ~96% | ~6 seconds | Development-phase value reached after overnight iteration |
| Later validation | 695/700, 99.3% | 0.9–5.4 seconds, ~10–14 seconds for the hardest case | Seven failure classes, each tested 100 times |
The 96% figure from development and the 99.3% from independent evaluation aren't the same measurement—the latter is a validation result from 700 trials testing seven failure classes, 100 times each. QuEra states that the five unsuccessful cases shared a common cause tied to conditions on the experimental apparatus side. There were no cases where the controller misjudged an incorrect frequency as a success. That said, both the testing and the root-cause analysis were conducted by QuEra itself; no independent replication, peer-reviewed paper, or complete test protocol has been published.
There's also data beyond the injected failures. QuEra ran a multi-week pilot in a normal, human-occupied lab environment and reports that all 43 naturally occurring mode hops were automatically recovered. The 700 controlled trials are complemented by fluctuations closer to real-world conditions. Still, this reflects performance on a dedicated testbed—QuEra identifies deployment to production QPUs as the next step.
Tuning 12 parameters without human intervention
If relocking addresses recovery from failure, tuning is about minimizing everyday drift. QuEra had Claude adjust 12 interdependent feedback parameters. While a human typically tunes values by visually tracking certain frequency bands, the agent recalculated the entire noise spectrum with every change. Across roughly 12,500 evaluations, QuEra reports zero incidents of the laser jumping to another mode.
After tuning, residual frequency noise dropped to one-fifth of its prior level. During roughly 19 hours of continuous unsupervised testing, there were zero lock losses, compared to 1.6 per hour before tuning. Furthermore, when compared against expert manual tuning using an independent measurement instrument untouched by the optimization system, performance was equivalent across the measured bandwidth. Notably, a resonance near 220 kHz that manual tuning had overlooked was suppressed by roughly a factor of 1,000, and a linewidth-equivalent metric narrowed from 58 kHz to 21 kHz.
When the same method was applied to a second wavelength used for quantum gates, the system found settings overnight and achieved a lock roughly 10 times quieter than the initial state, passing acceptance testing—made possible by reusing the sensitivity maps and safe operating ranges established during the first development effort. However, the tuning side still relies on Claude making decisions during execution. QuEra says it plans to consolidate this into a standalone tool similar to the recovery system going forward.
MHS is a connectivity layer that includes safety mechanisms
Claude's ability to read and operate different instruments was made possible through the Model Hardware Standard (MHS), an effort launched by Anthropic and the HHMI Janelia Research Campus. MHS drivers expose basic operations—like reading temperature or writing setpoints—in a common format. They also translate what a device can measure, what it can change, and how far it can safely go into machine-readable descriptions.
Agents can connect to these drivers via MCP, command line, or code APIs. For processes too fast for online inference or for long-running experiments, multiple instructions are bundled into a code file and run deterministically on the equipment side. QuEra's recovery script is an example of this pathway. Although it's a specification for connecting physical equipment to AI, it isn't designed to always hand control over to AI.
Safety isn't left entirely to model quality either. Operating ranges declared by the equipment, interlocks, and emergency stops are enforced at the hardware boundary, ensuring AI can only operate within those bounds. In this testing, a mechanism reverted to the previous state whenever a candidate setting caused a lock loss.
That said, MHS remains a limited research preview. Its specification and reference implementation aren't yet open source, and no independent safety evaluation has been published. It doesn't support devices lacking a programmable connection port, and Anthropic itself acknowledges that Claude's spatial and physical reasoning still requires expert oversight. To be judged as a truly model-agnostic common standard, it would need to be demonstrated that the same safety constraints hold with different models and third-party drivers.
Not the same as automating error correction
QuEra states that it has long automated recovery from common laser faults on Aquila, its 256-qubit analog machine available on Amazon Braket since 2022, citing a stated uptime above 99%. What this new effort targeted instead was rare, severe failures that fixed procedures struggled to handle and that still required expert judgment.
This distinction becomes more consequential for future machines. QuEra and AWS have outlined plans to bring "Libra," a system targeting a million logical operations across hundreds of logical qubits, to Amazon Braket in 2028. QuEra's roadmap goals include up to 256 logical qubits with a logical error rate of 10⁻⁶. As computations stretch from hours to days, QuEra explains, a mid-run laser outage could cause result corruption that error correction alone cannot rescue. Keeping laser experts stationed at remote sites around the clock also becomes harder to sustain as machine counts grow.
However, what's been disclosed here concerns laser control performance specifically. The 99.3% figure reflects relock success, not quantum computation success. Similarly, the fivefold noise reduction hasn't been directly measured against quantum gate fidelity, circuit completion rates, or logical error rates. Laser stabilization is an infrastructure condition that supports fault-tolerant quantum computing, but Claude has not automated quantum error correction itself.
Whether this translates into real gains for commercial deployment will depend on measuring mean time between failures and false-recovery rates once deployed on production QPUs, and on how much downtime and maintenance effort actually decrease. Only once improvements in laser noise are linked to measurable gains in quantum gate and logical qubit performance will the payoff of moving expert procedures into AI-generated code become visible as improved quantum computer uptime.