At MIT, AI Agents Pilot Routine Quantum Measurements

Within the EQuS group at MIT, agents powered by GPT-5.6 Sol and Codex are now orchestrating measurement and calibration sequences on superconducting chips. This setup saves time on repetitive tasks while still requiring human expertise when signals become weak or ambiguous.

Agents are regularly used for routine measurements

The EQuS group now regularly employs agents to manage routine measurements. The team produces a large number of standard chips, the characterization process of which can take several days of work for a researcher. By delegating these repetitive tasks to agents, scientists can focus on other activities. Thanks to Codex agents, experimental analysis and measurements can progress continuously without the need for constant supervision. However, experienced researchers are still able to determine calibration settings more quickly than current AI models allow.

A researcher organizes her work around agents and lab-connected code

Beatriz Yankelevich explains that she has set up an infrastructure guiding agents through measurement, theory, and chip design. She states that she can have multiple agents working in parallel on different problems, allowing her to spend most of her time interpreting results, designing new experiments, planning next steps, reading, and writing. The direct connection of agents to the lab enables code revision, testing on real measurements, and conducting long work sequences autonomously. For more innovative experiments, she assigns narrower objectives to the agents and leverages their ability to write, modify, and test control, analysis, and simulation code.

An experimental framework conducive to software automation

Once a superconducting chip is manufactured, packaged, and cooled, interaction occurs entirely through software, facilitating orchestration by agents. The connection of Codex to the lab software has allowed the AI to execute measurements, analyze results, and decide on the next steps. According to Beatriz Yankelevich, GPT-5.6 Sol has often autonomously completed routine measurement flows, saving her time and allowing experiments to proceed without constant oversight. She notes that this operation has freed up time for analysis, design, and planning future steps.

Results on an uncalibrated six-qubit chip

A test was conducted on an uncalibrated six-qubit chip, a standard format for evaluating the group's manufacturing process. Measurement skills describing the execution and evaluation of each experiment were provided to Codex. Using these skills along with the chip's design objectives, GPT-5.6 Sol determined parameters, controlled the equipment, and processed results, then adjusted measurements or saved data for future use. When signals were clear, a standard measurement sequence was completed with little intervention, including identifying transition frequencies, calibrating control and readout pulses, and estimating the quantum information retention time of a qubit.

Physical ambiguities and weak signals remain challenging for AI

Agents faced more obstacles when experimental signals were weak or contaminated by noise. In this context, the time required to find the right parameters increased, and the intervention of an experienced researcher sometimes proved essential. Observations indicate that current agents handle well-defined experimental flows effectively, but interpreting ambiguous physical results remains tricky. Calibrating qubits involves a series of interdependent measurements where each result guides the next, in a context where properties can drift and unexpected behaviors introduce inconsistencies. Seasoned operators know how to recognize and correct these evolutions, making this field a relevant use case for AI agents while also highlighting their limitations.

Why these qubits lend themselves to agent control

The EQuS group works on superconducting qubits cooled near absolute zero in dilution refrigerators. These devices, often referred to as artificial atoms, exhibit discrete energy levels manipulated by microwave pulses. Teams design and calibrate pulse sequences, then digitize and analyze responses to deduce resonance frequencies, quantum information conservation, and useful settings for computation. These systems, software-controlled and fabricated on-chip using known techniques, are well-suited for automated control. In a field where experimental preparation can take months and require hundreds to thousands of measurements, AI is positioned to assist, with the prospect that these approaches may one day enhance the simulation of complex materials and molecules.