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

What it claims

Two Purdue authors (Gusev, Bernal Neira) argue, in a perspective paper filed under cs.CY, that a very simple, repeatable agent-swarm procedure already produces theoretical research worth reviewing. In the parts they have checked, they found "no major scientific error", and several results are machine-checked in a proof assistant. They estimate a full review of the corpus would take them months. They state plainly that they do not claim all of it is correct or new. Most of it looks like the incremental work that makes up most of a scientist's output.

The argument does not depend only on frontier models. Some results came from GPT-5.6 Sol, which had been publicly available for almost two months. The paper's main contribution, by its own account, is Section 5: positions on review and trust, novelty when parts of the literature are closed to agents, publication, credit, the economic case for paying researchers rather than buying compute, budgets and fairness, training, and control.

How it works

The procedure is in Figure 1, with prompts reproduced verbatim in Appendix B. Appendix B covers literature collection, research, paper writing, and formal verification.

flowchart LR
  A[Scope: topic or field] --> C[Agent swarm]
  B[Papers + tools, may fetch more] --> C
  C --> D[Fresh agent reviews result]
  D --> E[Write up]
  E --> F{Verifiable?}
  F -- yes --> G[Proof or computation check]
  F -- no --> H[Left for human review]
  G --> C
  H --> C

The humans supplied no scientific ideas. The standing instruction was "make real, correct, useful progress, and do not stop." Every result is reviewed by a fresh agent, and formal verification is used where possible. The authors say this is neither the best nor the most cost-effective method. They chose it because it is simple. Their reasoning: if this already works, better methods will do more.

Evidence

• Output: Notes, paper-length drafts, proofs, and code in five areas of optimization theory and physical science, plus untested lab-experiment proposals in a sixth. Appendix C inventories the candidate papers and programs.
• Review backlog: The paper cites a survey of 637 US and UK scientists by Google, Google DeepMind and MIT FutureTech. Of the scientists who save time with AI, 46% spend more than a quarter of that saved time checking its output.
• Context results, cited rather than produced by the authors:
• Riemann zeta (Anthropic, August 2026): An unreleased Claude raised the share of zeros proved to lie on the critical line from about 42% (best result since 2020) to over 66%. The run used about 60 subagents and about 31M tokens. Humans condensed the argument and it was formalized in Lean.
• Navier–Stokes (OpenAI, 8 September 2026): About 10,000 agents produced, in 88 hours, a finite-time blowup proof for the forced 3D equations. Lean checking took a further 17 hours. The proof is still under independent review.

The Models, harnesses and cost section (4.2) is not in the fetched text, so no cost figures are given here.

How to use it

• The unedited corpus is released as of 25 September 2026.
• The authors invite readers to rerun the procedure in their own fields using the Appendix B prompts.
• They ask that failures be reported too, with the model and the date. Capability is uneven across tasks and changes quickly, so dated failures are the useful comparison point.

Limits

• Most of the corpus is unreviewed. Correctness and novelty are unknown for most of it.
• The lab experiments in the sixth area are proposed but not run.
• Agents could not access part of the literature, which weakens any novelty claims.
• The claims are scoped to the areas the authors tried, mostly theoretical work.
• The authors flag that their positions are tentative and expected to change. Section 3.3 lists which claims would not survive if the capability trend stops.
• The fetched text ends partway through Section 2.3. The detailed sections on output (4.3) and review evidence (4.4) were not available to this analysis.