What's this about?
OpenAI says coding agents can now complete well-defined research tasks that would take a skilled researcher a few days — its definition of an automated research intern. By mid-August, total use reached 3.1 agent-workdays for every human workday in the research organisation, but more than half of successful 4–8 hour tasks still involved human intervention.
What exactly is an “agent-workday”?
It is eight hours of agent runtime, not eight hours of proven useful work. Four agents running for two hours count as one agent-workday even if they duplicate effort or need correction. OpenAI uses the measure to show scale, while warning that runtime, code and experiments are only proxies for research progress.
Why is this worth noting?
This is rare operational data from inside a frontier lab: median researchers were spending more than $600 per day on agent inference at API prices, and experiment volume hit a record. It also shows the remaining bottleneck clearly — agents do more building and troubleshooting, while humans still choose priorities, judge results and steer long tasks.