OpenAI released a new report on agents in its own research workflows on November 18. For tasks estimated to require two to four hours of human work, its internal estimate of success without human intervention increased from 23% in March to 46% in October. The authors caution that researchers also change which tasks they delegate. The figures describe uneven progress, not the disappearance of the human role.

Why comparisons require care

Independent context comes from METR, but concerns a different organization. Its November 5 assessment examines AI R&D acceleration at Anthropic and stresses substantial uncertainty. Anthropic redacted much of the evidence METR considered most informative. That assessment illustrates the difficulty of evaluating the whole research process; it does not independently confirm the particular figures in OpenAI’s new report.

Our interpretation is that a task completed without intervention needs a definition before comparisons begin. If agents receive short, well-specified problems in one month and more complicated assignments in another, the raw percentage mixes several changes. Conversely, better task preparation could improve results even when the underlying model remains unchanged.

Preparation, execution and acceptance should therefore be separated. A person may select the inputs and establish an acceptance criterion before an agent works alone. A final human review may still be necessary. Without those distinctions, autonomous completion can easily acquire a broader meaning than the measurement supports.

From more activity to more knowledge

A hypothetical team could run twice as many experiments without obtaining twice as many useful findings. Some attempts might repeat an idea or fail later checks. This is our analytical example, not an additional statistic about OpenAI or Anthropic. It explains why output volume and research progress need separate measures.

For local tracking, we would record accepted results, reworked results and rejection reasons alongside the number of runs. A separate entry for human review would help show whether effort moved from writing to checking. The purpose is to understand completed work, rather than reward visible activity alone. A failed result should remain part of the accounting instead of disappearing from the success story.

What follows from the report

Publishing methods and limitations offers a more useful starting point than an undifferentiated claim of acceleration. The practical question is which parts of a workflow can now be delegated reliably and which still constrain the overall result. The answer may differ between coding, analysis and planning, even inside one organization.

SIExpo has not reproduced the internal evaluations. The news is the availability of more specific evidence about real laboratory use, with boundaries that readers can examine. It supports a discussion about the organization of research without establishing that AI already manages the entire scientific cycle on its own.