Signal AI Agent Swarms as Researchers: Progress, Challenges, and Open Questions

Summary

Researchers gave swarms of off-the-shelf coding agents only a scope statement, access to the scientific literature and computing tools, and a single standing instruction: make real, correct, useful progress and do not stop. Without being supplied any scientific ideas, the agents produced a large body of research notes, paper-length drafts, and formal proofs across five areas of optimization theory and physical science within weeks, and proposed untested laboratory experiments in a sixth area. The authors report finding no major scientific errors in what they have reviewed so far, and say several results have been proved using a proof assistant. Crucially, the agents produced output faster than the human authors could review it, with a full review estimated to take months. The paper argues research institutions are unprepared for this shift, since model capabilities are improving faster than institutional practices can adapt. It raises unresolved questions about trust in unreviewed results, what credit and publication should mean when the human contribution is a prompt, and how researchers can stay in control of work they cannot keep up with.