OpenAI ran 1,946 AI agents on a mouse brain atlas in 32 minutes

On 21 October 2026, OpenAI published a project that shows a different way to use AI in science. Together with the drug discovery startup PerturbAI and Scripps Research, the company took one very large biology experiment, cut it into 1,946 small pieces, and gave each piece to its own AI agent. All agents ran at the same time. Thirty two minutes later, the team had a structured report for every piece and a combined "reasoning atlas" of thousands of findings that scientists can now browse.

The work is described in a bioRxiv preprint by Alexander Sun, Joshua J. Park, David DeTomaso and colleagues, with corresponding authors at PerturbAI, OpenAI and Scripps. It has not been peer reviewed yet. Still, it is one of the clearest public tests so far of a simple question: if you have a mountain of data, is it better to give it to one strong AI agent, or to many agents that each look at a small part?

The short version

• The agents analyzed an existing whole brain in vivo Perturb-seq atlas from Scripps. In that experiment, researchers switched off close to two thousand genes in mouse brains and measured how millions of individual cells responded.
• The team ran 1,946 agents powered by GPT-Rosalind 5.5, OpenAI's life sciences model, one agent per gene disruption. The full run took 32 minutes.
• Without seeing the original paper, the agents matched or refined 48 of the 55 findings that could be drawn from the data alone, and partly recovered 4 more.
• At the largest workload, the distributed agents kept 711 times more literature references than a single agent doing the same job, which cited only 11.
• Thirty doctoral level biologists rated the reports blind. They preferred the distributed agent version in 75 percent of paired comparisons.
• The agents also pointed to new leads, including a difference between the genes Tsc1 and Tsc2 in the cerebellum and a possible link between the gene DCTN1 and cholesterol production in neurons.

The problem: experiments now outgrow the people who read them

Modern biology can produce data far faster than humans can interpret it. A CRISPR screen lets scientists knock out one gene per cell across a huge pool of cells. With Perturb-seq, they also read out the activity of every gene in each of those cells. The result is a table of how each gene disruption changes each cell type.

The Scripps atlas, published by Shi and colleagues in March 2026, did this inside living mouse brains. It covers close to two thousand perturbations across dozens of brain cell types. The authors of the new preprint, who include the team behind that atlas, are frank about the limit: even with experts in statistics, genomics and neuroscience, "substantial biology contained in the data remained under-interpreted." Standard pipelines produce lists of changed genes and pathway scores. Turning those lists into explanations means reading papers, comparing related genes and thinking about disease relevance, and that is slow work.

Sarah Teichmann, founding co-chair of the Human Cell Atlas, put the goal in one line in OpenAI's announcement: "We build cell atlases so that other scientists can make discoveries we haven't imagined yet."

How distributed biological reasoning works

The team calls its method distributed biological reasoning. The idea mirrors the structure of the experiment. One agent gets one gene disruption.

Each agent received the cell type resolved list of genes that changed after that disruption, a description of the experiment, a shared prompt and a coding environment with analysis tools and literature search. It could inspect the data, compare its gene with others and run its own small analyses. At the end, it wrote a structured report with findings, numbers, a biological interpretation, references, caveats and proposed follow up experiments.

A second stage then compared the reports with each other. It flagged how each finding related to published work, built a knowledge graph linking genes, pathways, cell types and evidence, and merged everything into 3,397 deduplicated relational findings, each with a priority score. This is the reasoning atlas that the team has now made public.

One design choice matters for trust. The original atlas paper appeared in March 2026, but GPT-Rosalind 5.5 has a stated knowledge cutoff of December 2025. The agents also did not get the manuscript or its conclusions. So when they "rediscovered" a result, they had to find it in the raw data, not remember it.

Test one: could the agents find what the human team found?

The authors listed 71 findings from the original study and checked each one against the agent reports. Of those, 35 were matched, 13 were refined (found and described in more detail), 9 were partly recovered and 14 were not covered.

The 14 misses are less worrying than they look. Some findings in the original paper relied on extra data that the agents never received. If you only count the 55 findings that could be drawn from the gene expression data alone, the agents matched or refined 48 of them, or 87.3 percent. With the four partial hits, they reached at least partial recovery on 52 of 55, or 94.5 percent.

This is a useful sanity check. Before you trust an AI system to find something new, you want to know it can find what is already known. Here it mostly did, without being told what to look for.

Test two: one big agent or many small ones?

The most interesting experiment in the paper is a controlled comparison. The team took the 893 gene disruptions that produced a clear response and gave them either to many agents or to one agent with a growing workload. The model, prompt, tools and data access stayed the same.

As the workload grew, the single agent did not scale. Its number of internal steps and tool calls leveled off, and it started skipping the literature. At the full workload of 893 disruptions, it cited only 11 references in total. The distributed agents kept 711 times more. OpenAI's own summary says "more than 700 references"; the preprint describes it as a 711 fold difference. Either way, the gap is huge.

There was also a quality signal. As its workload grew, more and more of the single agent's findings carried a "no literature retrieved" flag. The distributed agents kept roughly the same mix of flags at every scale. Among 1,519 high confidence findings, 27.6 percent agreed with published work, 41.9 percent were inferred from related work, 6.6 percent disagreed with it, and 23.8 percent had no precedent the agents could find.

The authors also tested whether these flags meant anything. They shuffled the gene labels so that each set of cell responses was attached to the wrong gene. The share of findings that "agreed" with the literature dropped from 29.3 percent to 13.2 percent, and to 6.1 percent when the labels came from genes outside the screen. That suggests the agents were really matching data to biology, not just writing plausible text.

The lesson goes beyond biology. Even a strong model has a limited attention budget in one session. Splitting a big job into many narrow jobs, each with a full budget, can beat one agent trying to do everything. Many teams building coding and research agents have reported the same pattern.

Test three: can the agents predict experiments they have not seen?

The team split the 1,946 targets into a training set of 1,674 and a held out set of 272. Agents saw the training data and had to predict how many genes would change after each held out disruption in one cell type, thalamic glutamatergic neurons.

They reached a balanced accuracy of 0.705, better than several simple baselines. They correctly called 82 percent of the cases with no effect (142 of 172) and 59 percent of the cases with a real effect (59 of 100). In some cases the reasoning was easy to follow: for the proteasome gene Psmc5, the agents predicted 520 changed genes, between two related genes in the training data, and the real number was 624.

This is not a breakthrough in prediction. Dedicated models do this kind of task too. But it shows the agents were using biological knowledge, not just summarizing.

Test four: what did human experts think?

To judge the value of the leads, the team recruited 30 doctoral level biologists. Each reviewed six reports, so every report got three independent reviews. Reviewers could search the web but could not use generative AI. They scored novelty, evidence, follow up priority and confidence on five point scales.

Two results stand out. First, the human scores followed the agents' own ranking: high ranked reports scored 3.14 on average, middle ranked 2.75 and low ranked 2.63. The agent and human scores correlated for novelty (rho 0.61) and follow up priority (rho 0.56). Second, in 60 paired judgments between distributed and single agent reports on the same topic, experts preferred the distributed version 45 times (75 percent), the single agent version 10 times, and called 5 a tie.

The leads: Tsc1, Tsc2 and a cholesterol clue

The paper highlights two leads that show what the system can add.

The first involves Tsc1 and Tsc2. These two genes work together to control mTOR, a key regulator of cell growth and metabolism. In most neurons, switching off either gene had similar effects, which fits their shared role. But in GABAergic neurons of the cerebellum, Tsc2 disruption changed 83 genes linked to synapses and calcium signaling, while Tsc1 disruption had almost no effect. The agents found no published precedent. This could matter for tuberous sclerosis, where cerebellar problems are strongly linked to TSC2 mutations and where learning and behavior symptoms remain hard to treat.

The second lead involves DCTN1, a gene whose mutations cause Perry syndrome, a rare neurodegenerative disease with Parkinson like symptoms. DCTN1 helps move cargo inside neurons. In the atlas, switching it off raised the activity of genes that make cholesterol. The agents connected this with research suggesting that when cholesterol transport fails, cells respond by making more. Blinded reviewers gave this lead a follow up priority of 4.3 out of 5. One reviewer wrote that the link "makes logical sense, and a quick Google Scholar search found nothing directly linking the two." Tomasz Nowakowski of UCSF, who was not an author, said the example shows how clues "scattered across different fields" can come together.

For contrast, a well known link between Apc and Wnt signaling scored high on evidence but low on novelty. The system did not only produce surprises; it also helped separate textbook results from fresh questions.

Where to be careful

The project is impressive, but a few points deserve care.

• The "months to a year" figure is an estimate, not a measurement. The authors built it from assumed times of 15 minutes per gene disruption and 10 minutes per paper read. Nobody timed a human team doing the same work.
• This is a preprint. It has not been peer reviewed, and several authors work for the companies involved.
• Leads are not discoveries. None of the new hypotheses has been tested in the lab yet. The authors say prospective experiments and independent datasets are needed.
• The agents can be wrong. The paper itself shows a case where agents predicted the gene Wdr26 would lower mTOR activity based on a study in a different cell type. Human review showed the opposite. The "disagree" flag helped catch it, but it is a reminder that review stays essential.
• One dataset, one model. The results come from one atlas and one model family. How well the approach transfers is still open.

Why it matters

Most public debate about AI in science focuses on one brilliant system solving one hard problem. This project points to something more practical. The bottleneck in many labs is not running experiments but reading them. A cheap, parallel layer of agents that writes a first pass report for every single result, with references and caveats, could change how scientists spend their time. Instead of reading thousands of rows, they review a ranked queue of leads.

The authors describe a loop: run a screen, let agents interpret it, use predictive models to suggest the next screen, and repeat. They are clear about the role of people in that loop. In their words, "Reading the literature, weighing explanations, and choosing decisive experiments would remain central to scientific training and judgment, while agents broaden the questions researchers can pursue."

Our take

The headline number, 32 minutes, is the least interesting part. Speed is easy when you can rent thousands of parallel model calls. The stronger result is the controlled comparison: with the same model and tools, many narrow agents read far more, flagged evidence more honestly and wrote reports that experts preferred. That is a design lesson anyone building agent systems can use today.

The real test comes next. If labs take the Tsc2 or DCTN1 leads to the bench and they hold up, distributed agents will move from an interesting demo to a normal part of the lab workflow. The data and the reasoning atlas are public, so other scientists can start checking right away.

Sources: OpenAI: Using thousands of AI agents to uncover new leads in biological data, bioRxiv preprint: Distributed AI agents enable comprehensive discovery of biological insights at scale, Whole brain atlas data on Hugging Face, UCSC cell browser, R&D World, OpenAI: Introducing GPT-Rosalind.