Thousands of agents read a brain atlas
OpenAI, PerturbAI and Scripps Research report that thousands of parallel AI agents, powered by OpenAI's GPT-Rosalind, analyzed an existing atlas of gene disruptions across millions of mouse brain cells in 32 minutes, combined their reports into a literature-linked "reasoning atlas" and proposed leads for brain disease. The preprint was posted on bioRxiv on 16 October 2026; OpenAI described the work on 21 October and made the atlas available.
Why it matters
The authors call the method distributed biological reasoning: one agent per gene disruption instead of one agent for the whole dataset. In their test, with the model and resources held fixed, a single agent cited 11 references at the largest workload against more than 700 for the distributed agents, and blinded human experts preferred the distributed interpretations. It is a way of spreading the reading of a large experiment across many agents; whether the leads it proposes (a cholesterol link in Dctn1, a pair of mTOR genes that differ in the cerebellum) survive experiment is not shown by the work.
What the sources say. OpenAI's post (21 October) and the preprint "Distributed AI agents enable comprehensive discovery of biological insights at scale" (posted 16 October 2026; corresponding author Xin Jin, Scripps) describe applying agents to a whole-brain Perturb-seq (CRISPR) atlas of mouse brain cells (the atlas itself came out in a manuscript of March 2026). Each agent investigated one gene disruption, analyzed cellular responses, searched the literature and proposed explanations; the reports were combined into a reasoning atlas. The agents used GPT-Rosalind-5.5, whose stated knowledge cutoff (December 2025) precedes the source manuscript, so the agents did not have that manuscript's conclusions. Parallel execution took 32 minutes; the team's illustrative estimate for comparable serial human analysis with literature review is months to a year. The post names leads: Tsc1 and Tsc2, which regulate mTOR, differ in the cerebellum; disrupting DCTN1 raises cholesterol-synthesis genes, raising a question about cholesterol distribution in neurons and neurodegeneration. The preprint's competing-interests statement says several authors are OpenAI employees and others are employees or co-founders of PerturbAI. Numbers. The 32 minutes are in both sources; "more than 700 references against 11" is in OpenAI's post, while the preprint's abstract says "100s times more references" and the figure 700 was not found in the preprint's text. The blinded-expert result is in both. What the record does not claim. It does not say the leads are confirmed biology: the post says the findings "help prioritize further experiments". The method was shown on one dataset. The study was written by authors some of whom work at OpenAI; no independent replication was read.