OpenAI deployed thousands of agents to decode brain cells in 32 minutes

Modern biology has a data bottleneck: researchers can disrupt thousands of genes across millions of cells, but connecting those dots with decades of scattered scientific papers usually takes months or a full year of grueling manual detective work.

Now, OpenAI, PerturbAI, and Scripps Research have tested a radically faster approach. By deploying thousands of distributed AI agents powered by a specialized model called GPT-Rosalind, the team analyzed a massive whole-brain CRISPR dataset from millions of mouse brain cells in just 32 minutes.

The swarm outperforms the solo model

Instead of dumping the entire dataset into one giant prompt, the researchers assigned individual agents to specific gene perturbations. Each agent analyzed cellular changes, combed through external literature, and generated biological hypotheses before combining them into a shared reasoning atlas.

The team ran a side-by-side test: as the workload grew, a single agent got overwhelmed and cited just 11 scientific papers, while the distributed swarm cited more than 700 references. In blinded evaluations, human experts consistently preferred the interpretations from the distributed setup.

Along the way, the agents surfaced genuine biological anomalies that humans easily overlook:

First, they caught an unexpected divergence between Tsc1 and Tsc2, two genes that regulate cell growth via the mTOR pathway. In the cerebellum, disrupting Tsc2 shifted dozens of genes tied to neuronal communication, whereas disrupting Tsc1 barely moved the needle. That discrepancy matches clinical observations in tuberous sclerosis, where cerebellar issues are heavily linked to TSC2 mutations.

Second, the agents flagged DCTN1, a gene linked to Perry syndrome (a neurodegenerative disease with Parkinson-like symptoms). When Dctn1 was disrupted, cholesterol synthesis genes spiked. The agents synthesized outside literature to hypothesize that impaired intracellular cholesterol transport might trigger this compensatory spike, accelerating neurodegeneration—a theory blinded experts deemed novel and worth immediate wet-lab follow-up.

Why it matters

We think this is one of the smartest demonstrations yet of how frontier AI should actually interact with science. It avoids the hype of claiming AI cures diseases overnight, and instead fixes a real operational choke point: the gap between generating massive experimental datasets and formulating testable hypotheses.

For researchers, the reasoning atlas is available now. It does not replace the wet lab—every hypothesis still requires physical bench validation—but turning months of literature synthesis into half an hour shifts the bottleneck directly to experimental execution.

It also provides an architectural blueprint for AI builders. When dealing with complex reasoning across millions of data points, throwing a single monolithic model at the problem leads to degradation and laziness. Orchestrating specialized parallel agents that cross-reference external literature proves far more reliable.

The clear risk, however, is scientific hallucination baked into literature biases. If published papers contain uncorrected errors, an agent swarm might weave those flaws into highly convincing, plausible-sounding dead ends. Whether this setup truly accelerates drug discovery will depend entirely on how many of these hypotheses survive contact with actual bench experiments.