OpenAI's thousands-of-agents biology run read a mouse-brain CRISPR atlas in 32 minutes, but the evidence is a preprint and human experts preferred it 75% of the time, not always
One agent per gene disruption, 700+ literature references versus 11 for a single agent, and a Dctn1-to-cholesterol lead. Verdict: the fan-out pattern is worth copying; the biology is still untested in the lab.
The research team estimated that comparable human analysis and literature review, conducted one perturbation at a time, would take months to a year. The agents reproduced findings from human analyses while uncovering additional leads.
Comprehensive take
The durable finding is not a cholesterol hypothesis, it is the control experiment: with model and budget fixed, one agent per question beat one agent for everything on literature coverage by roughly 65 to 1. Copy the pattern, verify the biology in the lab.
Use cases
Research-agent builders
Fan out one agent per entity (gene, account, document) with its own search budget, then merge, instead of one long-context agent.
Computational biologists
Compare related genes and unexpected responses in the published reasoning atlas to shortlist experiments.
Evaluation teams
Reuse the design: hold model and compute fixed, scale the workload, and measure citation coverage and blinded expert preference.
Specs
| Run time | 32 minutes over data from millions of mouse brain cells source |
|---|---|
| Human baseline | Months to a year, a team estimate source |
| Literature coverage at largest workload | 700+ references (distributed) vs 11 (single agent) source |
| Blinded expert preference | Distributed reports preferred 75% of the time source |
| Preprint | bioRxiv 2026.10.15.759436v1, posted 2026-10-18, not peer reviewed source |
Insights
- Axion reading: the 11 versus 700+ reference gap measures search coverage, not correctness; neither number says how many citations were relevant.
- Axion note: the 75% preference figure is OpenAI-reported; sample size and judging protocol were not stated in the post.
What OpenAI actually did
OpenAI, PerturbAI and Scripps Research pointed thousands of agents, powered by a model OpenAI calls GPT-Rosalind, at an existing whole-brain CRISPR atlas: a dataset showing how disrupting individual genes changes cells across the mouse brain. Each agent took one gene disruption, analysed the cellular response, searched the literature and proposed a mechanism. The reports were then merged into what the team calls a reasoning atlas. OpenAI says the run covered data from millions of mouse brain cells and took 32 minutes. The team's own estimate for a human doing the same one perturbation at a time is months to a year; that figure is an estimate, not a measured baseline.
The accompanying bioRxiv preprint, posted on October 18, frames this as distributed biological reasoning. Its abstract adds claims the blog post only gestures at: without access to the source manuscript the agents recovered its conclusions, and they predicted held-out perturbation effects.
The number that matters for builders
The most transferable result is the single-agent control. The team held the model and resources fixed and grew the workload. The single agent searched less of the literature as the task got bigger, citing 11 references at the largest workload, while the distributed agents cited more than 700. In blinded comparisons, human experts preferred the distributed reports 75% of the time. That is a clear majority, and it also means a quarter of comparisons did not go that way. OpenAI does not state how many experts judged, how many reports were compared or how ties were handled in the post, so treat 75% as a headline to check against the preprint, not as a settled effect size.
The lesson is architectural. A long-context single agent tends to satisfy itself with the strongest signals. Splitting the problem so that every unit of work gets its own search budget forces coverage. If you are building research agents over any large tabular or screening dataset, that is the part to borrow.
The leads, and how far to trust them
The agents flagged a Tsc1 versus Tsc2 difference in the cerebellum: disrupting Tsc2 altered dozens of genes involved in neuronal communication there, while Tsc1 produced little change. They also connected Dctn1, a gene linked to Perry syndrome, with raised expression of cholesterol-synthesis genes, and asked whether cholesterol distribution inside neurons contributes to neurodegeneration. Human experts rated the Dctn1 hypothesis novel, consequential and worth following up, according to OpenAI.
Note the verbs. These are hypotheses that prioritise experiments; the post says outright that further experiments are needed to test the proposed mechanisms. No wet-lab validation is reported in the material OpenAI published, and the work is mouse data, not human. The post is also vendor-authored, with OpenAI as one of three participating organisations and no independent replication yet.
Who should care
Computational biologists can already inspect the reasoning atlas, which OpenAI says is on GitHub, and the underlying atlas is on Hugging Face. Agent engineers should read the preprint for the fan-out and merge design. Anyone evaluating GPT-Rosalind itself should wait: the post gives no model specifications, pricing or access terms in the section we reviewed.