The overnight shift no one volunteered for
Beatriz Yankelevich, a grad student in MIT's Engineering Quantum Systems Group, has found someone—or something—willing to babysit superconducting qubits at 3 AM. She connected GPT-5.6 Sol through Codex to her lab's dilution refrigerator and let it run the tedious, months-long process of calibrating quantum computing experiments while she worked on other things. Or slept. Or fabricated chips in the cleanroom.
This is not a demo. This is production experimental physics, the kind where a single qubit chip can take days to characterize and where getting the measurements wrong means starting over. And it's working well enough that EQuS now regularly uses agents for routine measurements.
The technical details here matter because they reveal both what current AI agents can actually do in the wild and where they still fall short. Spoiler: they're great at the boring stuff, less great when the data gets weird.
What calibrating a qubit actually involves
Superconducting qubits are artificial atoms cooled to near absolute zero. You manipulate them with microwave pulses, read out their state, and hope they hold quantum information long enough to be useful. Before you can do any real quantum computation, you need to calibrate everything: resonance frequencies, pulse shapes, readout fidelity, coherence times.
This calibration is a cascade of interdependent measurements. Each result shapes what you try next. The workflow is well-defined but adaptive—you need to recognize when a signal is clean versus noisy, adjust parameters on the fly, and decide whether to refine the current measurement or move on.
Traditionally, this meant a researcher sitting at a terminal for hours or days, running measurements, eyeballing plots, tweaking parameters, and repeating. It's the kind of work that demands expertise but doesn't always reward it with insight. Perfect agent territory.
Codex as lab infrastructure
Yankelevich's setup is straightforward but non-trivial. She gave Codex measurement-specific skills—essentially functions explaining how to run and evaluate each experiment type. She provided the chip's design targets. Then GPT-5.6 Sol chose measurement parameters, operated the hardware, analyzed the data, and either refined the measurement or saved the result for the next step.
The key architectural choice here is tight integration. Codex isn't generating Python scripts for a human to review and run. It's directly controlling lab hardware, reading sensor data, making decisions, and iterating. The agent is in the loop, not outside it.
When signals were clear, the system completed standard measurement sequences with minimal human intervention. It identified transition frequencies, calibrated control and readout pulses, and measured coherence times. Routine chip characterization that normally takes days now runs overnight or while Yankelevich is in the cleanroom.
"I can have agents running measurements for many hours overnight or while I'm working in the cleanroom. I can check in from my phone, see what they've done, and steer them if something needs fixing or if I want to explore a different direction."
Where the wheels come off
GPT-5.6 Sol struggled when experimental signals were weak or noisy—exactly the situations where an experienced physicist's intuition matters most. The agent took longer to converge on suitable parameters and sometimes needed guidance.
This failure mode is instructive. Current frontier models are excellent at executing well-defined procedures and decent at iterative refinement within known bounds. They're much worse at interpreting ambiguous physical data or recognizing when something fundamentally unexpected is happening.
The paper doesn't sugarcoat this. EQuS uses agents for routine measurements but still relies on researchers for novel experiments and edge cases. That's the right engineering trade-off: automate the 80% that's tractable, reserve human cycles for the 20% that requires deep understanding.
Multi-agent workflows for multi-domain research
The more interesting development is Yankelevich's move beyond single-task agents. She's built infrastructure to guide agents through measurement, theory, and chip design simultaneously. Multiple agents work on different problems in parallel while she focuses on higher-level work: interpreting results, devising experiments, planning next steps, reading, writing.
This is the 2025 agent story in microcosm. Single-model architectures are for demos. Multi-agent systems are for shipping. The complexity is real—coordinating agents, managing handoffs, ensuring consistency—but so are the productivity gains.
The killer feature isn't that agents can write pulse calibration code faster than humans. It's that they can run that code against real hardware, collect data, refine the approach, and iterate for hours without supervision. The feedback loop closes at machine speed.
What this tells us about agent readiness
This work landed at an interesting moment. We've spent 2024 debating whether AI agents are overhyped vaporware or the future of work. The EQuS deployment suggests a more nuanced answer: agents are production-ready for structured experimental workflows with clear success criteria and tight feedback loops.
Three factors made this work:
- Software-mediated interaction: Once the qubit chip is cooled, everything happens through code. No robot arms, no physical manipulation, just API calls.
- Well-defined measurement protocols: Qubit calibration is complex but not ambiguous. There's an established sequence, known techniques, and quantitative success metrics.
- Tolerance for iteration: Physics experiments expect to run many trials. An agent that takes 10 attempts to converge is slower than an expert but still useful if it runs unattended.
Remove any of those factors and the value proposition gets shakier. Agents aren't replacing domain expertise; they're amplifying it by handling execution overhead.
The gap between routine and novel
The clearest limitation is the routine-versus-novel divide. GPT-5.6 Sol excels at standard chip characterization because the workflow is repeatable and the success criteria are clear. For novel experiments, Yankelevich gives agents narrower goals and leans more heavily on their code-writing abilities.
This tracks with what we're seeing across agent deployments in 2025. Current models are excellent at automating known-good procedures and decent at exploratory work within bounded domains. They're not yet capable of the kind of open-ended scientific creativity that leads to breakthrough insights.
That's fine. Most research isn't breakthroughs. Most research is careful, systematic exploration of parameter spaces and validation of hypotheses. If agents can handle that 80%, they're incredibly valuable.
The overnight experiment as product category
There's a product insight lurking here: the overnight experiment as a service. Industries with long-running tests and iterative refinement—materials science, drug discovery, manufacturing optimization—could benefit from agents that run experiments while humans sleep.
The economic case is compelling. If an agent can compress three days of researcher time into 24 hours of unattended runtime, the value is obvious. You're not just saving labor; you're accelerating the entire research cycle.
The challenge is generalization. Yankelevich built custom infrastructure for her specific experimental setup. Turning this into a general-purpose platform means abstracting the measurement primitives, standardizing the data formats, and handling the long tail of edge cases across different experimental domains.
That's a hard engineering problem, but it's an engineering problem, not a research problem. The core capability exists. Someone just needs to productize it.
What's next
EQuS is now using agents routinely for chip characterization. That shift from "interesting experiment" to "part of the workflow" is the milestone that matters. Production adoption by researchers with better things to do than babysit calibration scripts is the strongest possible endorsement.
The next frontier is handling the weird cases—weak signals, unexpected resonances, hardware glitches—where current models still need human guidance. Improvements in multimodal reasoning and uncertainty quantification will help. So will better tools for researchers to quickly course-correct agents mid-run.
But even without those improvements, the current system is useful enough to ship. Yankelevich checks in from her phone, steers the agents when needed, and spends most of her time on high-level work. That's the future: not agents replacing researchers, but agents handling execution so researchers can focus on the thinking that actually requires a PhD.
The qubits are getting calibrated either way. Might as well let the robots stay up late.