Michael JohnsonOn October 21, 2026, OpenAI described a study in which thousands of AI agents analyzed data on interventions affecting millions of mouse brain cells in 32 minutes. Instead of using one long prompt, the team divided the work among agents: each examined a separate genetic intervention, compared cellular changes with publications, and proposed a possible explanation. The result was an open catalog connecting genes, cellular effects, and scientific literature.
OpenAI, PerturbAI, and Scripps Research took part in the project. Their study description and article preprint present a practical use case for agent systems: reviewing a broad range of experimental results in parallel and turning identified connections into questions for laboratory testing. This is especially useful when there is more data than a research team can manually compare with scattered literature.
How the analysis was divided
GPT-Rosalind agents worked with a whole-brain atlas showing how disruptions to individual genes affect different cell types. Each agent examined its own intervention, studied the cellular response, searched for relevant scientific publications, and formulated a hypothesis. The separate reports were then combined into a reasoning atlas—a catalog linking observations to possible mechanisms and sources.
According to the research team, comparable manual analysis, one intervention at a time, would take several months to a year. This comparison applies to the described volume of work and the team's estimate; the agents reanalyzed data that had already been collected rather than conducting new biological experiments. The resulting hypotheses help identify which observations merit further investigation.
Why the team used thousands of agents
OpenAI compared the distributed approach with a single agent while keeping the model and available resources the same. As the task grew, the single agent cited 11 sources in the largest scenario, while the distributed system cited more than 700. In a blind comparison, specialists preferred reports from distributed agents to those from a single agent 75% of the time.
These figures reflect a specific experiment and expert evaluation. For research teams, the practical takeaway is to assess not only the speed of analysis, but also source coverage, citation quality, and whether the results become clear, testable hypotheses.
From an unexpected connection to the next experiment
One example concerns the DCTN1 gene, whose mutations cause Perry syndrome, a neurodegenerative disease with symptoms resembling Parkinson's disease. In the atlas data, disruption of Dctn1 was accompanied by increased activity in genes associated with cholesterol synthesis. By comparing this observation with publications on cholesterol transport, the agents posed a question: could disruption of DCTN1 alter cholesterol distribution within neurons and affect their damage?
Another hypothesis emerged from comparing Tsc1 and Tsc2. Although both genes are involved in regulating mTOR, disruption of Tsc2 in certain cerebellar cells changed the activity of dozens of genes associated with neuronal communication; the team observed a much smaller effect in that region for Tsc1. Such results help narrow the search: researchers can select specific cells and mechanisms for subsequent laboratory testing.
The published reasoning atlas catalog lets researchers explore connections and choose directions for further work. The project's main practical value lies in the scale of its initial review: the system quickly connects large experimental datasets with literature and proposes ideas to test. Researchers can use these ideas to decide which hypotheses to take into the laboratory.