Agentic AI
OpenAI's AI agents multiply workload, not efficiency—even its data proves it
OpenAI's push to deploy autonomous research agents has paradoxically increased the supervision burden on human researchers, offsetting gains in automation.
The New Stack
OpenAI deployed what it calls an "automated research intern"—an agent capable of completing structured research tasks that would ordinarily consume weeks of human work. Yet the company's own internal metrics reveal a counterintuitive outcome: the more agents researchers deploy, the more oversight they must provide.
By August 2026, OpenAI's research organization was running agents at a ratio of 3.1 agent-workdays per human workday, with compute costs soaring—the median researcher spent over $600 daily on inference, while high-volume users exceeded $7,000. But counting agent-hours masks the real constraint: human attention.
Agent productivity versus oversight burden#
A crucial distinction separates agent-hours from completed work. While researchers can dispatch multiple agents simultaneously, each produces a stream of code, experiment results, and technical decisions requiring review. Running more agents doesn't reduce human workload; it multiplies the tasks waiting for human validation.
OpenAI's definition of success is deliberately modest. The "research intern" must complete well-structured tasks that would otherwise take a skilled person several days—yet a human remains the ultimate decision-maker. The company's next milestone, an "automated AI researcher" with genuine autonomy, isn't targeted until March 2028.
Where agents contribute—and where they stall#
Using a framework from Epoch AI, OpenAI tracked agent activity across six research categories: Decide, Design, Build, Run, Analyze, and Communicate. Between January and August, activity climbed across all six areas, though agents remained reluctant to tackle the hardest phase—deciding which research problems to pursue in the first place.
Agents excelled at mechanical work: writing code, tracking experiment runs, and providing technical support so effective that debugging office hours saw attendance plummet at some teams. Yet volume doesn't equal value. OpenAI noted that code output and experiment counts are easy to measure but reveal nothing about research progress. Compute usage spiked as experiments multiplied, but whether agents accelerated discovery remains unclear.
When OpenAI assessed agent performance across varying task complexities, agents showed improving success rates between January and July—yet humans had to intervene on more than half of the completed tasks that would have taken four to eight hours to complete manually.
