Swarms of agents unleashed on five theoretical domains produce more results than their authors can reread

Two researchers gave swarms of code agents a perimeter, access to literature and an instruction: to progress without stopping. In a few weeks: drafts of articles and formal proofs, in five areas.

Sergey Gusev and David E. Bernal Neira conducted a deliberately stripped-down experiment: entrusting swarms of commercial code agents with a simple statement of scope—sometimes a narrow subject, sometimes an entire field—access to literature and computational tools, and then a permanent instruction: to make real, correct, and useful progress, and not to stop. No scientific ideas have been provided by humans.

In a few weeks, the agents produced a large volume of research notes, paper-length drafts, and formal proofs in five areas of optimization theory and the physical sciences, and proposed laboratory experiments not yet performed in a sixth. The authors are careful not to say that everything is right or new, but insist that this is not noise. In what they checked, they found no major scientific errors, and several results are demonstrated in a proof assistant.

Proofreading becomes the scarce resource

The most uncomfortable point of the paper is a reversal: agents produce faster than their authors can proofread, a complete review being estimated to take several months. Hence a series of open-ended questions on the trust placed in unreviewed results, on the meaning of signing and counting publications when human input is limited to an instruction, and on the reason for institutions to pay researchers rather than buy computing time. The authors believe that research institutions are not ready, publish the gross production of agents as of September 25, 2026, and invite everyone to repeat the experience in their own field.