No. 032

Several new platforms now offer researchers AI agents for specific tasks, from manuscript formatting to physics simulation to bioinformatics code rewriting. Two independent preprints document what LLMs do to citation rhetoric and peer-review quality when they handle those tasks at scale: citation patterns become less critical and popular-biased, and review recommendations vary by a factor of six across models on the same papers.

Agents in Research

• Distributed AI agents enable comprehensive discovery of biological insights at scale

  bioRxiv, October 16 2026

  Thousands of parallel agents, each analyzing one part of a whole-brain Perturb-seq atlas, completed in minutes what the authors estimate would take months of human analysis, and blinded human experts preferred the distributed-agent interpretations over those of a single agent in a head-to-head comparison.

• Agentic AI systems for scientific discovery

  Nature Reviews Methods Primers, October 22 2026

  A 19-author methods primer covering how AI agents work across literature search, simulation, and automated experiments also addresses how to evaluate these systems, keep workflows reproducible, and keep human judgment involved, making it a useful reference point for researchers designing agent-assisted pipelines.

Agent Tools and Platforms

• Why choose Google Antigravity for science

  Google, October 2026

  Google is offering 10,000 scientists six months of free access to its top Antigravity tier and the Gemini 4 Argon model, with a Science plugin connecting to more than 30 life-science databases including UniProt, PubChem, ChEMBL, and the AlphaGenome Atlas; applications opened October 13 and the underlying science skills are open source.

• Lemma by Axiomatic AI

  Axiomatic AI, September 2026

  An AI research assistant for physics and engineering entered open beta, offering notebook-based workflows for simulation (Tidy3D, MEEP, QuTiP, Qiskit), dimensional and symbolic checks, Lean 4 formal proofs, and LaTeX manuscript drafting, with demo claims that have no independent verification.

• Academic Paper Formatting Skills for Claude

  GitHub, October 21 2026

  Three free Claude skills by researcher Jiao Yu prepare manuscripts for submission to JAMA Network, Gerontological Society, and 20 sociology and demography journals, covering journal-specific word counts, reporting-guideline mapping (STROBE, CONSORT), mock peer review, and reviewer-response table drafting.

• Rewriting Scientific Software with Agents

  AI Exchange @ UVA, October 23 2026

  Stephen Turner surveys AI agents rewriting bioinformatics tools into Rust, including a case where one tool's processing time dropped from 15 hours to 15 minutes on a single dataset (a figure from an OpenAI field report, cited as such), and describes Claude Code's code-modernization plugin, which validates old and new versions against the same inputs.

Peer Review and Citation

• Major Revision, Minor Accountability: LLMs as Reviewers of Published Articles in Top-Tier Journals

  arXiv cs.DL, October 22 2026

  Four LLMs reviewed 200 CC BY papers already published in Nature, Science, and Cell; 75.6% of reviews recommended major revision or rejection, with individual model rates ranging from 17.5% to 100%, and when a single field-specific flaw was planted in each paper, detection rates ranged from 17.0% to 60.5%, making LLM review a low-sensitivity and highly model-dependent signal.

• Citing Less Critically: LLMs Reshape the Rhetoric and Reach of Scientific Citation

  arXiv cs.DL, accepted EMNLP 2026

  Six LLMs filling masked citation contexts across 1,746 NLP conference papers cited less critically than the human authors, favored popular and older papers over recent niche work, and drew on authors socially distant from the citing scholar rather than from within the coauthor network.

Mathematics and AI

• Mathematics, artificial intelligence and the opportunities ahead

  Academy for the Mathematical Sciences, October 22 2026

  Terry Lyons argues that AI solving hard problems faster does not reduce the need for mathematicians, because the value of mathematics lies in new representations and understanding rather than in answers, and that publishing and evaluation systems must adapt to results produced at speed, with early-career researcher support as a priority rather than a casualty.

• AIAS+ 2026: Chen Institute Symposium for AI Advancing Science and Society

  Chen Institute, November 5-7 2026

  An annual symposium organized by Jennifer Chayes, Jennifer Doudna, Omar Yaghi, and others focuses on where AI for science falls short: what an ideal AI scientist would need, how to handle AI discoveries that humans can verify but not fully understand, and how to protect the scientific record as AI-generated research grows, with sessions featuring Yejin Choi, Terence Tao, and Dashun Wang.