936 AI agents run for 12 weeks, spontaneously invent peer-review journal and research norms
ProfBuehlerMIT · x · 2026-11-21
A 12-week exploratory study deployed 936 autonomous AI agents across 35 human-operated labs (Stanford, MIT, ETH Zurich, UCSF and more) on a shared preprint server and messaging workspace.
Key findings:
- Agents produced 936 studies with full code and data, exchanging 9,500 messages across 229 self-created chat rooms
- Unprompted, they debated study designs, publicly retracted mistaken claims, acknowledged negative results, and founded an agent-run peer-review journal (Journal of Agent Research) after catching other agents cheating in code
- Agent-run reproducibility audits confirmed findings matched experimental output; agents also publicly corrected flaws in human papers
- In blinded evaluations, agent papers scored 9.02/10 on experimental rigor vs 7.00/10 for human preprints, though humans still won on novelty and significance
The authors released the full 12-week archive of manuscripts, code, data, reviews and chat transcripts, documenting how scientific norms and structures emerge in large-scale agent societies.
More from AGI Musings
- Mathematician claims he and Claude proved the Riemann Hypothesis, pending verification — YiMaTweets · 2026-11-21
- Terence Tao: Riemann Hypothesis needs something beyond current AI, maybe human-AI by 2050 — ChrisGPT · 2026-11-21
- Build for automation that serves humans: staying in control — gdechichi · 2026-11-21
- Multiplayer coding agents: agent attention is unbounded, human attention isn't — manosaie · 2026-11-21
- Konrad Kording doubles down: AI people should hold put options as a hedge — KordingLab · 2026-11-21
- The bitter lesson, interpretability edition: humans steer only the weakest AI in the stack — burny_tech · 2026-11-21