OpenAI's AI Agents Generate More Tasks Than They Solve, Internal Data Reveals

OpenAI achieved a milestone it announced last fall: deploying what it calls an "automated research intern," an agent designed to tackle clearly scoped research assignments that would ordinarily consume multiple days of a researcher's time. Throughout 2026, the use of coding agents expanded steadily across the organization. By mid-August, these agents had accumulated 3.1 agent-workdays for every single human workday spent by researchers. The financial toll climbed alongside this activity—the typical researcher spent more than $600 daily on inference costs at API rates, while top-tier users in the 90th percentile exceeded $7,000 per day.

Agent hours versus useful output

Yet an agent-workday and a human workday measure fundamentally different things. OpenAI standardizes agent labor into eight-hour equivalents, but this metric only captures duration—not deliverables. Since researchers can operate multiple agents simultaneously, the numbers reflect how long agents are active, not necessarily what they accomplish. For engineering teams, this creates a new burden: deploying more agents simultaneously can multiply the work being performed, but it also multiplies the work humans must oversee.

OpenAI's definition of a research intern remains deliberately narrow: the agent must complete well-defined research tasks requiring several days of skilled human effort, with a human retaining ultimate control. The company has set its sights on a more ambitious target—an automated AI researcher—which it aims to achieve by March 2028.

The company's next goal, an automated AI researcher, is one they hope to reach by March 2028.

Supervision becomes the constraint

Using a framework developed by Epoch AI, OpenAI categorized agent activities into six domains—Decide, Design, Build, Run, Analyze, and Communicate—and observed growth across all categories between January and August. Notably, agents contributed minimally to the Decide phase, which involves determining which research directions to pursue. Much of their work remained operational: writing code for research and infrastructure, tracking experimental progress, and delivering technical assistance so effective that some teams discontinued their debugging office hours entirely.

Increased agent activity does not automatically translate to meaningful research advancement. OpenAI acknowledges that code volume and experiment counts are straightforward to measure but reveal little about actual progress. As experiments multiplied, computational requirements surged proportionally. When OpenAI evaluated agent performance across tasks of different complexity levels, it discovered that despite rising success rates from January through July, human intervention remained necessary for more than half of successful tasks that would have required four to eight hours of human work.

Security incidents limit Astra deployment

Once engineers can orchestrate multiple agents that spawn their own subagents, the supervisory challenge intensifies dramatically—tracking which runs have diverged, examining code changes, and determining what is production-ready or suitable for training pipelines. Astra's ability to maintain persistent agent sessions and accept multi-day research assignments has made this oversight problem worse, not better. OpenAI recognizes that as agents assume greater responsibility for execution, the remaining research work—the aspects most resistant to automation—will demand proportionally more of an engineer's attention, creating a practical ceiling on how much agent output a single person can reasonably review.

On July 20, a cascade of agent-driven outages severely damaged OpenAI's research infrastructure, forcing the company to shut down its training container service and later restore it with stricter controls. Nearly four weeks later, on August 7, OpenAI imposed additional restrictions after preliminary analysis indicated Astra could reach the "Critical" level under its Preparedness Framework, confining the model to higher-security research domains and implementing additional protections that developers have begun experiencing as unexpected API interruptions.

Workloads shift between models fast

GPU allocation for Astra-class systems dropped 59.2% in the week following the restrictions, yet the freed compute resources did not remain idle. Researchers redistributed much of the workload to alternative models, which saw GPU allocation climb 17.2% and offset roughly 85% of the Astra reduction. Rather than curtailing overall work volume, the constraints simply redirected it elsewhere, demonstrating how readily workloads migrate when one system faces restrictions.

OpenAI's research teams are delegating increasingly complex assignments to agents, running greater numbers concurrently, and launching more experimental trials. Yet translating this activity surge into measurable research velocity remains elusive—and OpenAI continues to grapple with how to structure pricing around these shifting patterns.

OpenAI's researchers are handing off larger jobs to agents, running more of them at once and launching more experiments, but whether that translates into faster research is harder to measure — and OpenAI is still figuring out how to price it.

Source: The New Stack