Over twelve weeks, 936 autonomous AI agents in 35 human-run laboratories published 936 studies, each with its code and data, on a shared preprint server, according to A Society of AI Scientists, a paper submitted on 19 November.
The authors, from institutions including MIT and Stanford, say the agents also started their own peer-review venue, the Journal of Agent Research.
Important: not investment advice
This article is for information and education only. It is not investment, financial, legal or tax advice, and it is not a recommendation or an offer to buy or sell any security, ETF, option or other financial product. FourWeekMBA is not a broker-dealer or a registered investment adviser. IPO shares, leveraged and inverse ETFs, and options carry significant risk, including the loss of your entire investment; leveraged and inverse ETFs reset daily and can lose value quickly. Prices, availability, fees and eligibility change and can differ by country and broker, so check the prospectus and your broker’s terms. Figures marked “our arithmetic” are our own calculations from the cited filings and may contain errors. Do your own research and consider speaking to a licensed financial professional before making any investment decision.
Business Pill · The Model and the Harness
An evergreen Business Engineer explainer on the difference between an AI model and the harness around it. It teaches the general idea only and does not discuss this study.
The key insight: In twelve weeks, 936 AI agents in 35 labs published 936 code-backed studies and founded their own review journal; LLM judges rated them 9.02 vs 7.00 on rigor against human preprints, while humans won on novelty.
What the Agents Did
The study ran from 19 August to 12 November 2026. Agents exchanged 9,500 messages across 229 chat rooms they created themselves, the paper says.
Humans provided compute, set budgets and chose broad themes. The agents chose the specific hypotheses, designs and analyses, according to the authors.
Co-author Samuel Schmidgall wrote that “no human wrote code or text” in the agents’ papers, in a post announcing the study.
The cohort ran on hardware from CPU-only laptops to AMD GPUs and B200 clusters. Topics ranged from language-model fine-tuning to mechanistic interpretability, continual learning and solar physics.


The Norms That Emerged
Agents publicly withdrew mistaken claims before publishing. Across the 19 published cases of self-correction, every correction came before upload, with a median lead of 21.4 hours, the paper reports.
Negative results were cited as often as positive ones. Three of the eight most-cited studies reported negative findings, according to the authors.
Agents in one laboratory founded the Journal of Agent Research. Its agent editors inspected code and scrutinised experiment results, the paper says.
Rigor Versus Novelty
Two LLM judges rated 202 matched pairs of agent and human papers with authors and affiliations removed. Agent papers scored 9.02 on experimental rigor against 7.00 for human preprints, on a 10-point scale.
That gap is 2.02 points (our arithmetic; the paper reports +2.01, with a 95% interval of +1.74 to +2.29). Human papers scored higher on novelty and significance.
Earlier this week OpenAI said its agents now complete 46% of two-to-four-hour research tasks, as we reported. This study tests something different: whether many independent agents can check and build on each other’s work.
The Structural Read
As we read the study, the agents were strongest where research is checkable: code, data and reruns. They were weakest at choosing which questions matter.
As we read the authors’ caveats, the bottleneck shifts from running experiments to judging which results are worth anyone’s attention.
The authors, A Society of AI Scientists
“Agents still lack the kind of independent scientific thinking that is required for great research.”
Three Implications
FOR RESEARCH LABS Agents produced code-backed studies at volume, but the authors say humans still had to pick the noteworthy ones.
FOR PUBLISHERS AND REVIEWERS An agent-run journal inspected code, and audits caught errors that text review missed, the authors say.
FOR AI COMPANIES The paper rates agent work higher on rigor and lower on novelty than recent human preprints.
What Is Not Established
The authors write that agents “still lack the kind of independent scientific thinking that is required for great research”. They often gravitated toward familiar topics and small variations on existing benchmarks.
Making sense of the output took “substantial human effort”, and the authors say interpreting the significance of results will remain human.
Much of the work used small models and short experiments. Comparisons across the study’s growth stages are observational, not randomized, so the paper says scale cannot be shown to have caused the changes.
The ratings came from AI judges, not human reviewers. The authors have released the full archive of manuscripts, code, data, reviews and chat transcripts, and opened the agent preprint server to the public.
The Bottom Line
In this study, 936 agents produced 936 code-backed studies in twelve weeks and built their own review culture; AI judges rated them more rigorous than recent human preprints but less novel, and the authors say choosing worthwhile questions remained a human job.
Important: not investment advice
This article is for information and education only. It is not investment, financial, legal or tax advice, and it is not a recommendation or an offer to buy or sell any security, ETF, option or other financial product. FourWeekMBA is not a broker-dealer or a registered investment adviser. IPO shares, leveraged and inverse ETFs, and options carry significant risk, including the loss of your entire investment; leveraged and inverse ETFs reset daily and can lose value quickly. Prices, availability, fees and eligibility change and can differ by country and broker, so check the prospectus and your broker’s terms. Figures marked “our arithmetic” are our own calculations from the cited filings and may contain errors. Do your own research and consider speaking to a licensed financial professional before making any investment decision.
97,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.
A note on sourcing. All figures on the study come from the paper’s alphaXiv page, its abstract and summary, and the authors’ announcement thread; Schmidgall’s quote is from his X post. The 2.02-point gap is our arithmetic from the reported 9.02 and 7.00; the paper reports the difference as +2.01. The rigor and novelty ratings were produced by two LLM judges, as the paper states.
Sources: Ahmad, Anderson, Qu et al. (incl. Schmidgall, Moor, Buehler), A Society of AI Scientists, alphaXiv, submitted 19 Nov 2026, and authors’ announcement thread · Samuel Schmidgall on X, 19 Nov 2026 · FourWeekMBA, OpenAI agents finish 46% of 2-4 hour research tasks, 18 Nov 2026









