A whole-brain CRISPR atlas that would have taken a research team months to sift took 32 minutes when OpenAI, PerturbAI and Scripps Research handed the work to close to 2,000 agents running on GPT-Rosalind, the lab-focused model OpenAI published this week alongside a bioRxiv preprint.
The dataset covers millions of mouse cells. Each agent took one gene perturbation, read how the cells responded, searched the literature for an explanation and wrote a report.
What the swarm surfaced
Two findings stand out. Disabling Tsc2 shifted dozens of genes governing nerve-cell signaling in the cerebellum, while its close relative Tsc1 barely moved anything, a gap that could matter for tuberous sclerosis. Shutting off Dctn1, tied to Perry syndrome, raised activity in cholesterol-making genes, pointing toward an experiment no one had queued.
The agents also cited more than 700 references at their largest workload. A single agent managed 11.
Why splitting the work mattered
In blinded comparisons, outside experts preferred reports from the distributed setup over a single agent’s output 75 percent of the time. OpenAI has posted a Reasoning Atlas linking genes, cell effects and literature on GitHub.
The caveat is the one that trails every agent demo: these are hypotheses. Labs still have to run the assays. What changed is how quickly a hypothesis now reaches the bench.
