OpenAI says coding agents are changing how its researchers work

OpenAI says coding agents are moving from occasional assistants to a regular part of research work. In a September 6 report, the company describes researchers using agents throughout the day, often in concurrent sessions, while contributing code faster and running more experiments. The report says that, by mid-August, the median researcher in its research organization was using more than $600 per day of inference at API prices, and that the organization was using the equivalent of 3.1 agent-workdays for every human workday.

Those numbers are internal measurements, not a promise that every team will see the same results. OpenAI also cautions that faster coding and more experiments do not automatically mean that the entire research process advances at the same rate. People still set priorities, judge results, and decide whether work should be scaled, paused, or deployed.

A September 8 Codex case study shows what that workflow looks like outside ordinary application code. MIT researcher Beatriz Yankelevich connected Codex to laboratory software for superconducting-qubit experiments. The agent could run measurements, analyze results, and choose what to try next; the researcher supplied measurement-specific skills and the chip's design targets.

The boundary is important: the case study reports that the agent handled clearly defined, routine workflows well, but needed guidance when signals were weak or noisy. That is a useful pattern for software teams too. Agents are strongest when the task has explicit steps, observable outputs, and a human who can intervene when the situation becomes ambiguous.

The practical takeaway is less “let the agent run everything” and more “turn repeatable work into a supervised loop.” Define the task, give the agent the domain context it needs, make intermediate results inspectable, and keep a person responsible for decisions that depend on judgment. Teams should also measure the cost of the loop and the quality of its outputs, not just the number of generated changes. A faster agent that creates more review or debugging work may not improve the whole system.

The latest examples suggest that agentic development is becoming a workflow-design problem as much as a model-selection problem for many research teams.