OpenAI just published the most useful public data I've seen on what AI agents actually accomplish.
The headline: as of mid-August 2026, roughly 3.1 agent-days of work run in parallel for every eight-hour day a researcher works. The median researcher burns $600+/day in agent inference.
The number buried at the end: more than half of those tasks still required human intervention at least once.
The task list is unglamorous and that's the point — writing code, setting up training environments, running evals, debugging failures, monitoring runs. Every item has a clear finish line. "Decide what to research" is still human.
The detail I keep coming back to: some teams cancelled their standing technical office hours because agents now fix enough of the environment breakage themselves. They automated the interruptions, not the interesting work.
OpenAI calls this an "automated research intern" and targets a full "automated researcher" by March 2028. That gap — intern vs colleague — is exactly the "half needed a human" number.
For anyone delegating real work to agents, the takeaway is boring but true: delegate bounded tasks with a finish line, keep the judgment, and budget for review time.
More, including Pachocki's "An Alien Mind" essay on failing CoT oversight: https://toolkitcreators.com/guides/ai-agent-reality-check-openai-data
Top comments (0)