Quantum computing promises to revolutionize material science, but the path there is paved with tedious, repetitive labor. Before a single complex algorithm runs, researchers must calibrate superconducting qubits through hundreds of preliminary measurements. This process can consume months of a scientist’s time. Beatriz Yankelevich, a graduate student in MIT’s Engineering Quantum Systems Group (EQuS), decided to offload this grunt work to an AI agent. By connecting GPT-5.6 Sol to her laboratory’s control software via Codex, she created a system that could run, analyze, and refine experiments on quantum chips with minimal human oversight.
The setup leverages the fact that modern quantum experiments are entirely software-driven once the hardware is cooled. Yankelevich provided the AI with specific skills for running and evaluating measurements on a standard six-qubit chip. The agent didn’t just follow a script; it chose parameters, operated the hardware, and interpreted the resulting data to decide its next move. When signal quality was high, GPT-5.6 Sol successfully identified transition frequencies and calibrated pulse sequences autonomously. This capability transforms the researcher’s role from a constant monitor to a strategic supervisor. “I can have agents running measurements for many hours overnight,” Yankelevich noted, adding that she can steer the process remotely if issues arise.
However, the technology isn’t flawless. The agent struggled when faced with weak or noisy signals, often requiring guidance from an experienced human to interpret ambiguous physical results. This highlights a critical boundary in current AI capabilities: while agents excel at defined workflows, they lack the intuitive problem-solving skills of seasoned physicists when data gets messy. Yet, even with these limitations, the time savings are substantial. EQuS now uses these agents regularly to characterize chips that previously took days to process manually.
This integration suggests a hybrid future for lab work rather than full automation. AI handles the volume and consistency of routine calibration, freeing scientists to focus on experimental design and high-level analysis. As quantum devices scale up, the ability to delegate these foundational tasks will likely become a prerequisite for progress, not just a convenience. The bottleneck shifts from data collection to data interpretation, forcing labs to rethink how they train the next generation of quantum engineers.
