936 AI Agents Wrote 936 Studies in 12-Week Study
A paper titled A Society of AI Scientists appeared on alphaXiv on November 18, 2026, with an update stamped November 21, and describes a 12-week exploratory run in which 936 autonomous agents operated across 35 human-operated laboratories inside a shared preprint and messaging workspace (alphaXiv). The abstract states the run produced 9,500 messages, 229 agent-created chat rooms and 936 published studies, including code and data (alphaXiv).
Why this matters
For anyone building, buying or deploying agents, the headline figures are the clearest signal yet of scale: a four-figure population of autonomous agents ran continuously for three months inside a workspace designed to mimic how human labs share preprints and coordinate, and it generated output at roughly one published study per agent. That ratio — and the fact the agents organized themselves into hundreds of chat rooms without apparent central direction — is the part of the record that is actually verified. Commentary circulating online has gone further, describing agent editors reviewing code and experiments, agents outperforming or losing to human papers on specific criteria, and benchmark-style evaluation results (FourWeekMBA). None of those claims can be confirmed from the available primary-source excerpt, and readers evaluating this as evidence of agent research capability should treat them as unverified.
What the abstract confirms
The authors' own words are narrow: "Here, we present an exploratory twelve-week study in which 936 autonomous agents across 35 human-operated laboratories interacted in a shared preprint and messaging workspace" (alphaXiv). That sentence establishes the cohort size, the lab count, the duration and the setting. It does not name the models used, specify a benchmark, state an evaluative sample size, or identify the authors' institutions — claims attributing the work to MIT, Stanford, ETH Zurich or UCSF are not supported by the paper record itself.
What remains unconfirmed
No limitations section, peer-review status, or quality assessment of the 936 studies is available in the current record. Whether a human or automated process judged the studies' validity, whether the "journal" framing reflects a formal review process, and whether any of the 936 outputs have been checked against existing literature are all open questions the abstract does not answer. Until a fuller version of the paper, author list, or methods section surfaces, the verified story is limited to the activity counts: 936 agents, 35 labs, 12 weeks, 9,500 messages, 229 rooms, 936 studies.


