OpenAI says a Codex agent ran a six-qubit calibration loop at MIT

OpenAI says the model was harnessed through Codex ^[1], and in the demonstration it selected parameters, ran the relevant measurements through the lab hardware, assessed the results and chose subsequent actions with minimal human intervention ^[5]. For that loop to close, the instrument control layer has to be reachable from wherever the agent executes code, and every step has to return something the next step can read. The lab has to supply both of those.

The steps named are concrete: identify qubit transition frequencies, calibrate control and readout pulses, estimate coherence ^[4]. Three task types ^[13], each ending in a number, each feeding the next. OpenAI describes calibration as interdependent measurements that have to be performed repeatedly, with the system interpreting data before it proceeds ^[6].

The write-up lists three prerequisites for the pattern: access to the relevant tools, a workflow the agent can understand, and boundaries for when it should stop or ask for help ^[10]. The third one takes work: a stop condition means somebody has written down what a bad measurement looks like in terms a program can evaluate. OpenAI says noisy or ambiguous data may still need human guidance ^[7], and the case study does not establish that AI can reliably manage every experimental condition without oversight ^[18].

On transfer, dev.to argues that many businesses run work with the same shape: gather data, apply a known procedure, inspect the result, determine the next action ^[11], and it offers recurring test runs with established parameters and quality-control checks that produce structured measurements as candidates ^[17]. Qubit calibration supplies two things for free that an office workflow often will not. The procedure already exists in written form, and it is run again and again ^[6]. Point an agent at a procedure nobody has written down and it will be inventing the procedure. Inventing it is a different experiment from this one.

The record here is OpenAI's own case study on Codex and quantum computing experiments, relayed by dev.to ^[15]. How often Yankelevich had to step in, and how the agent's calibration results compared with the lab's manual runs, are not in the write-up ^[14]. It does narrow its own claim, to a substantial portion of a routine workflow and not the full scientific process of developing quantum computing research ^[8]. I do not often see a vendor case study narrow its own claim. The benefit it reports is time: more of the researcher's for experiment design, data analysis and planning ^[9].