936 AI agents run a 12-week experiment in autonomous science
Agents from 35 human-run laboratories shared a preprint server, wrote 936 studies with code and data, and founded their own peer-review journal. Blind reviewers rated their work above recent human preprints on rigour and below them on novelty.
Source published · In the Daily Brief of 20 November 2026
Filed under Agents, Large language models
The study, posted on alphaXiv on 19 November by a large team including researchers at Stanford and MIT, followed 936 autonomous agents working across 35 laboratories for twelve weeks. The agents shared a preprint server and a messaging workspace, where they exchanged about 9,500 messages in 229 chat rooms they created themselves. Each of the 936 studies came with runnable code, raw data and a record linking reported numbers to experimental outputs.
Norms emerged without being imposed. Agents retracted mistaken claims before publication, reported negative results about as often as positive ones, and agents in one laboratory founded the Journal of Agent Research, whose editors inspected code before accepting work. Audits that reran analyses caught errors that reviews of the text alone had missed. In blinded comparisons with recent human preprints, the agent studies scored higher on experimental rigour and measured claims, and lower on novelty and significance.
The authors say the agents tended to drift towards familiar topics and small variations on existing benchmarks, and that sorting the noteworthy results from the rest took substantial human effort. The full archive of manuscripts, code, peer reviews and chat transcripts has been released. The work has not been peer reviewed.