Nearly 1,000 AI agents ran a 12-week research community, publishing 936 papers
SRSchmidgall · x · 2026-11-20
Across 35 labs (Stanford, MIT, ETH Zurich, UCSF and more), nearly 1,000 AI agents autonomously conducted research for 12 weeks on a shared server — no human wrote code or text — producing 936 papers with full code and data.
Key findings:
- Agents unpromptedly debated study designs in chat, publicly retracted their own mistakes, found real errors in human papers, and founded their own peer-review journal that caught other agents cheating in code
- In blind reviews against recent human arXiv preprints, humans still won on original ideas; agents won on rigor (9.0 vs 7.0)
- Some agents independently reproduced results similar to later human papers, including failure modes in self-distillation and feature reproducibility in mech interp
"The scientific community is about to get a lot bigger, and most of the new members won't be human."
More from coding & agent
- Self-hacking with abliterated models: biggest risks were infra side channels, not app code — cyb3rops · 2026-11-20
- Pocket City: a pixel-art Amsterdam canal simulation built entirely on real geodata — anselm · 2026-11-20
- Mnemos vs Claude Code memory test details: 65%/41%/42% recall across three blind tests — RileyRalmuto · 2026-11-20
- Solo-built Mnemos memory system recalls 2x more facts than Claude Code's native memory — RileyRalmuto · 2026-11-20
- Kaggle and Eigent launch AgenticRL-Bench to test coding agents on real RL incidents — dejavucoder · 2026-11-20
- Suno MCP now lets you generate original songs inside ChatGPT and Claude — heyshrutimishra · 2026-11-20