OpenAI's internal coding agents are now handling a measurable share of the company's own AI research workload. Researchers are offloading tasks including literature review, code generation, experiment setup, and result analysis to agents running autonomously for hours at a time. The company is tracking this shift with internal metrics on agent usage rates, experiment velocity, and task complexity, and the early numbers show a meaningful compression in the time between hypothesis and result.
The more interesting story is not the productivity gain headline. It is the structural change in how research is organized. When agents can run parallel experiments overnight without human supervision, the bottleneck moves from execution to evaluation. Researchers are spending more time deciding what to test and interpreting outputs, less time writing boilerplate and waiting. The piece includes specific data on how experiment volume has changed and what categories of tasks agents handle most reliably versus where they still fail.
This report is worth reading in full because it is a rare case of a lab publishing operational data about its own internal tooling rather than a product announcement. The methodology for measuring research acceleration, what counts as agent-assisted versus agent-completed, is itself a contested and underexamined problem. OpenAI is making early choices about how to define and count this work, and those definitions will matter as the rest of the field tries to benchmark similar systems.
[READ ORIGINAL →]