936 AI agents just published their own research. Was any of it good?
A 936-agent research community wrote and posted its own preprints, and the verdict on quality, not the volume, decides whether this is a milestone or a noise machine.
Nine hundred and thirty-six AI agents ran an experiment in which each one carried a research project from question to finished paper, publishing end to end on a preprint server of their own while exchanging thousands of messages. The open question of whether the resulting science is any good is now the central problem for anyone betting on autonomous research pipelines: generation has become cheap, and the value is migrating to whoever can grade the output.
Nine hundred and thirty-six AI agents spent the run exchanging thousands of messages, and every one of them published a complete piece of research, end to end, on a preprint server of their own. A Science reporter got an exclusive seat inside the whole thing, eavesdropping on a research community with no humans in it. The question hanging over every one of those papers is blunt, and it is the only question that matters: was any of it good?
That question is not rhetorical. It is the reason this experiment deserves more attention than any individual paper it produced. Preprint servers have historically been the human front door of science: researchers post a manuscript before peer review so the field can read it, argue with it, and cite it in near real time. Here, the authors, the future reviewers, and the server itself were all agents. If the output holds up, this is a milestone. If it does not, it is a very efficient way to manufacture scientific noise at a volume no journal, and no reviewer, can absorb.
Start with the scale, because the scale is the story. Nine hundred and thirty-six agents. Thousands of messages. Each agent carrying its own project from a starting question to a finished write-up. For context, a productive academic lab might hold a dozen people and publish a handful of papers in a year. The experiment folded something that normally requires a field's worth of institutions, shared standards, editors, review cycles, and years of elapsed time, into a single run. Coordination, not raw intelligence, is usually the expensive part of research. This run made coordination nearly free.
"End to end" is doing heavy lifting in that sentence. In a normal pipeline, humans frame the question, design the method, run the analysis, write the manuscript, and then defend it against people who want to find the error. The agents here owned the entire chain, and they published into infrastructure they ran themselves. That means the usual checkpoints, an editor, a referee, a skeptical rival reading the methods section at midnight, were simply not in the room.
So, was it any good? That is the open question the experiment leaves behind, and it is the correct one to ask. Quality in research is not a vibe. It is whether the claims are true, the methods reproducible, the data real, and whether anyone would catch a mistake before it propagates. Volume measures throughput. It says nothing about whether a paper would survive contact with a hostile reader. A thousand papers nobody can reproduce are not a thousand contributions. They are a thousand liabilities with formatting.
If agent-authored preprints become routine, the pressure moves to provenance and evaluation. Science already has a quantity problem: publication counts are a proxy for productivity, and proxies get gamed by humans every day. Hand hundreds of agents a preprint server and you industrialize the gaming. The scarce resource stops being the draft and becomes the verdict. Who ran this, on what data, with what checks, and who is accountable when it is wrong? Those questions are about to get much harder to answer, and much more expensive to ignore.
For executives well outside the lab, this is a preview rather than a curiosity. The architecture on display, many narrow agents, one shared protocol, high message volume, a single finished artifact, is the same shape as due diligence, equity research, code review, and legal discovery once you automate them. The teams that come out ahead will not be the ones that generate the most drafts fastest. They will be the ones that built the grading function first, the internal mechanism that decides which machine output gets trusted, funded, or shipped.
Watch three things from here. Who audits the 936. Whether those papers get cited or quietly ignored. And whether the scientific community treats agent-authored preprints as contributions or as spam. That is what separates a clever demo from the first draft of a new way of doing research. The reporter got to eavesdrop on it. Everyone else will find out through the citation record.
This story's Key Insights and Take-aways are locked.
Create a free account to unlock Executive Actions for one credit.
Register to UnlockAlways free for Executives Club members. Join the Club
More in Science
Anthropic's I.P.O. Made Rank-and-File Researchers Worth Tens of Millions
Researchers who joined just a few years ago are now worth tens of millions, and the windfall resets the pay benchmark for every private A.I. lab still paying in paper.
Sea Lion Culling Strategy Fails: New Model Says Killing Them Won't Save Salmon
The latest research from AAAS overturn a long-held conservation assumption, forcing policymakers to rethink critical resource allocation in marine ecosystems.
Microplastics do more than pile up: they react with water to damage cells
Scientists have found that plastic particles can generate highly reactive molecules in water, a mechanism that could explain links to heart attacks, strokes, and neurological disorders.


