Thousands of GPT-Rosalind agents mine a brain cell atlas

What happened?

On October 21, OpenAI described (announcement) joint work with PerturbAI and Scripps Research. Thousands of agents powered by GPT-Rosalind analyzed existing data from a whole-brain mouse CRISPR atlas, which shows how gene disruptions affect millions of cells. The run took 32 minutes; the team estimates comparable human analysis would take months to a year.

OpenAI calls the approach distributed biological reasoning. Each agent took one gene disruption, analyzed the cellular response, searched the literature and proposed an explanation. The reports were combined into a searchable reasoning atlas linking genes, cellular effects, publications and possible mechanisms.

What it found

• The agents reproduced findings from human analyses and surfaced additional leads.
• Tsc1 and Tsc2 behaved differently in the cerebellum: disrupting Tsc2 changed dozens of genes involved in neuronal communication, while Tsc1 had little effect. That may matter for tuberous sclerosis research.
• Disrupting Dctn1 raised expression of cholesterol-synthesis genes, leading to a testable hypothesis linking cholesterol distribution and neurodegeneration.
• In blinded comparisons, experts preferred distributed-agent reports over single-agent reports 75% of the time.

Why it matters

For developers the most useful part is the scaling test. With the same model and resources, a single agent cited only 11 references at the largest workload, while the distributed agents cited more than 700. Splitting the work did not just speed things up; it changed quality, because one agent with a huge context simply looks at less.

The lesson carries beyond biology: when a large dataset breaks into many independent pieces, fanning out to many small agents and merging the results often beats stuffing everything into one long context.

What to watch

These are hypotheses, not validated results; the proposed mechanisms still need lab testing. The reasoning atlas is available to researchers now. If you build agentic pipelines, the 11 versus 700 references result is a solid argument for the fan-out pattern.