On October 21, 2026, OpenAI published “Using thousands of AI agents to uncover new leads in biological data”, describing work with PerturbAI and Scripps Research, with an accompanying bioRxiv preprint. The team split the analysis of an existing whole-brain CRISPR atlas, a dataset of how gene disruptions affect cell types across millions of mouse brain cells, among thousands of agents powered by GPT-Rosalind, an approach OpenAI calls distributed biological reasoning. Each agent took one gene disruption, analyzed cellular responses, searched the literature and proposed explanations; the reports were merged into a public “reasoning atlas”. The run took 32 minutes; the team estimates comparable one-at-a-time human analysis would take months to a year.
OpenAI reports the agents reproduced findings from human analyses and that their judgments of novelty and value aligned with blinded human experts. Highlighted leads include a difference between Tsc1 and Tsc2, two genes that jointly regulate mTOR: in the cerebellum, Tsc2 disruption altered dozens of genes involved in neuronal communication while Tsc1 disruption changed little, which OpenAI connects to the link between TSC2 mutations and cerebellar abnormalities in tuberous sclerosis. Another is that disrupting Dctn1, whose mutations cause the Parkinson-like Perry syndrome, increased expression of cholesterol-synthesis genes, prompting a testable hypothesis about cholesterol distribution in neurons and neurodegeneration.
In a fixed-resource comparison, a single agent given the whole dataset searched less as the workload grew, citing 11 references at the largest workload against more than 700 for the distributed agents, and blinded experts preferred the distributed-agent reports 75% of the time.
Why it matters: it is a concrete demonstration of scaling out many parallel agents over one large dataset, rather than one long-running agent, and of how much more literature coverage that buys.
What it does not show: these are hypotheses, not validated discoveries; none of the highlighted leads is reported as tested experimentally, and the speed comparison with human analysis is the team’s estimate rather than a measured head-to-head.