35 labs, 936 AI agents publish research and peer-review each other
Understand this story
In an exploratory deployment, 936 AI agents from 35 human-run laboratories published 936 studies with accompanying code and data between August 19 and November 12, 2026, and exchanged 9,500 messages in 229 agent-created chat rooms. Humans remained responsible for providing computing resources, setting budgets, and selecting broad topics, while the agents themselves chose specific hypotheses, designs, and analysis methods. The research team states that comparisons across different stages of development are observational analyses, not randomized causal tests; the recorded review objections are execution-verified objections in journals, not a random audit of all papers in the entire cohort; and the reviews are blind text evaluations by LLMs against scoring rubrics, not human expert peer review, with reviewers unable to run code or inspect figures.
AI overview based on source reports · Updated 42 min ago
Coverage timeline
Follow the coverage to explore different perspectives.
- Hacker News AIA Society of AI Scientists
This article describes an exploratory deployment in which 936 AI agents across 35 human-run labs shared a research archive and message space, publishing 936 studies with code and data and exchanging 9,500 messages in 229 agent-created chat rooms between August 19 and November 12, 2026. Agents chose their own hypotheses and methods while humans supplied compute, budgets and broad topics, and the staged rollout was observational rather than randomized. The strongest evidence came from comparing review modes: text-only reviewers flagged 10 of 16 revised drafts that had already disclosed limitations, while scripted checks run on uploaded data overturned key conclusions in 77 of 156 major objections (49.4%) and 24 of 120 minor ones (20.0%).
Activity for this story
Not enough continuous observations to show a trend yet.