OpenAI just showed off one of the more unusual jobs its models have taken on: running quantum computing experiments. In a new labs writeup, OpenAI details how an MIT researcher pairs GPT-5.6 Sol with Codex to autonomously run experiments, analyze the results, and calibrate qubits. This is a real research loop, not a demo. And the same playbook works for anyone who wants an AI agent to drive a technical workflow instead of just answering questions.
What stands out here is the shift from chat to action. The model isn’t suggesting what to do next. It’s doing it, checking the output, and adjusting.
Quick Start
You’ll learn how to turn a capable model into a hands-on lab assistant that runs code, reads the results, and tunes the system on its own. To follow this pattern you need three things: a coding agent (OpenAI uses GPT-5.6 Sol with Codex), programmatic access to whatever you’re controlling (in this case lab hardware and its APIs), and a clear definition of what a good result looks like.
- Run the experiment autonomously. The researcher hands the agent the experiment and lets it execute. Why this matters: experiments in a physics lab involve dozens of small, repetitive setup and run steps. Letting the model handle the execution frees the researcher to think about the science instead of babysitting scripts. Codex is the piece that lets the model actually write and run the code that talks to the equipment.
- Analyze the results. After a run, the agent reads the output and interprets it. This is the step most automation skips, and it’s the one that makes the whole thing useful. Raw data from an experiment means nothing until someone decides whether it worked, what the numbers imply, and what to try next. GPT-5.6 Sol does that reading and reasoning, then feeds the conclusion back into the loop.
- Calibrate the qubits. Qubits, the basic units of a quantum computer, drift and need constant tuning to behave correctly. The agent uses what it learned from the analysis to recalibrate them, then runs again. That closes the loop: run, read, adjust, repeat, with the model steering each pass. According to OpenAI, this is where the setup earns its keep, because calibration is tedious, frequent, and exactly the kind of work that eats a researcher’s day.
Why this is a big deal
Quantum labs are one of the least forgiving places to test an AI agent. The hardware is finicky, the feedback is noisy, and a wrong move wastes expensive machine time. If a model can hold its own here, it can hold its own in a lot of ordinary engineering and data work too.
There’s a broader signal in this. Over the past year the frontier labs have pushed models from writing code to running it, checking it, and acting on the result. OpenAI’s quantum example is that trend showing up in a hard science setting rather than a software one.
A few things to keep in mind
- Give the agent a clear success signal. It can only calibrate toward a target you’ve actually defined.
- Keep a human on the outcomes, not the keystrokes. The point is to review conclusions, not approve every command.
- Expect the loop to matter more than any single run. The value comes from run, analyze, adjust repeating fast.
Next steps beyond the tutorial
You don’t need a quantum lab to use this pattern. Pick a workflow you run over and over, where you can measure success and give an agent programmatic control: a data pipeline, a testing suite, a model-tuning job, an ops task. Start by letting the agent run one step and report back. Add analysis next. Only then hand it the controls to adjust and re-run on its own.
The move from AI that talks to AI that operates is well underway. OpenAI’s full writeup on the MIT quantum work is worth a read for the specifics of how they wired it together.