OpenAI calls its agents research interns. Measure accepted output, not runtime.

What changed

OpenAI published an account on 6 September of how coding agents are used inside its research organisation. The company says it has reached the goal it announced last year of an automated research intern that can complete well-defined research tasks under human direction, including work that could take a skilled researcher several days.

OpenAI reports that by mid-August the organisation used 3.1 agent-workdays of effort for every human workday. The measure converts agent runtime into eight-hour units and includes parallel sessions and downstream agents. It shows a large increase in machine work being attempted alongside researchers.

The ratio should not be read as a 3.1 times productivity result. Agent runtime can include failed experiments, duplicated paths and work that requires correction. Independent analysis of the disclosure also notes that more than half of successful tasks lasting four to eight hours still involved human intervention.

• The milestone is defined by OpenAI and measured inside its own research organisation.
• Agent-workdays quantify runtime, not accepted scientific output.
• Humans still define tasks, steer longer work and judge results.

The operating lesson for businesses

The important change is managerial. One person can supervise several agent workflows at once, which alters how research, analysis and engineering work is queued. The bottleneck can move from producing a first attempt to specifying the task, monitoring progress, reviewing evidence and deciding which result to keep.

That shift can create hidden cost. OpenAI reports high inference use among its researchers when valued at API prices. A business that measures only employee time saved may miss model spend, retries, review time, security controls and the cost of plausible but incorrect work entering a decision process.

Research teams also need a clear acceptance boundary. Agents can search, code, run experiments and summarise, but senior staff remain accountable for methods, evidence and conclusions. A concurrency target without quality gates can increase the volume of work that reviewers must reject.

What businesses should do next

Pilot the model on a narrow class of tasks with an objective acceptance test. Suitable examples include reproducing an analysis, preparing a literature map, running a bounded data check or building a prototype that must pass a known test suite. Avoid beginning with open-ended decisions where correctness is hard to observe.

Track attempted work and accepted work separately. For each task, record agent runtime, compute cost, human steering, review minutes, defects, evidence quality and whether the output changed a decision or shipped result. This turns an impressive concurrency measure into business unit economics.

Set a review capacity limit before adding more parallel agents. The system is working when accepted output rises without an unsafe increase in reviewer load or unresolved errors. Keep humans responsible for research questions, risk judgments and final claims until the organisation has evidence for a narrower delegation boundary.

• Measure acceptance rate and reviewer effort alongside runtime.
• Price failed paths and retries into the workflow.
• Use objective checks for the first delegated task class.
• Scale concurrency only when review capacity keeps pace.