GPT-5.6 Sol agents automate routine superconducting-qubit calibration

OpenAI says an MIT researcher has connected GPT-5.6 Sol, working through Codex, to laboratory software used for superconducting-qubit experiments. In a case study published by OpenAI News, graduate student Beatriz Yankelevich describes using the agent to operate instruments, analyze measurements, and select the next calibration step.

The test involved an uncalibrated six-qubit chip from MIT’s Engineering Quantum Systems Group. With measurement-specific skills and the chip’s design targets, the agent identified transition frequencies, tuned control and readout pulses, and measured how long qubits retained quantum information. When signals were clear, it could complete standard workflows with little researcher input.

That matters for AI-tool builders because the system was not simply generating code in isolation. Codex was connected to the software layer controlling real hardware, creating a loop in which the model measured results, interpreted them, adjusted parameters, and continued the experiment. Yankelevich says agents can now run for hours overnight or while she works elsewhere, with remote check-ins from her phone.

The limits are equally important. Weak or noisy signals slowed the agent and sometimes required an experienced researcher’s intervention. The results point to a practical boundary for autonomous scientific workflows: well-defined, repeatable procedures are within reach, while ambiguous physical behavior still benefits from human judgment.

EQuS says it now regularly uses agents for routine chip measurements. By handling characterization that can take researchers several days per chip, the tools leave more time for experiment design, analysis, theory, and chip development. The broader pattern is a useful blueprint for builders: connect agents to trusted domain-specific skills, live instrumentation, and checkpoints rather than treating autonomy as a standalone chatbot feature.

Source: OpenAI News