Thousands of agents, one perturbation each: distributed reasoning over a brain CRISPR atlas

OpenAI, PerturbAI and Scripps Research published a case study this week describing how they ran thousands of AI agents in parallel over a whole-brain CRISPR atlas (perturbation data from millions of mouse brain cells) and turned it into a set of candidate research leads in 32 minutes. It is not a new model; it is a worked example of an orchestration pattern.

What they actually did

The dataset records how disrupting individual genes changes the behaviour of different cell types across the mouse brain. The team split the analysis so that each agent, running on GPT-Rosalind, owned a single gene disruption. Each agent analysed the cellular response, searched the literature, proposed an explanation, and wrote a report. The reports were then merged into what they call a reasoning atlas: a linked map of genes, cellular effects, publications and candidate mechanisms. OpenAI calls this distributed biological reasoning, and has released the atlas for other researchers.

The team estimates that a comparable human analysis, done one perturbation at a time with a proper literature review, would take months to a year. According to the post, the agents reproduced findings from earlier human analyses, and their own ratings of novelty and scientific value matched those of blinded human experts.

The leads

The first is a split between Tsc1 and Tsc2, two genes that jointly regulate mTOR. Disrupting either one looked similar in several neuron groups, but in the cerebellum only Tsc2 disruption moved dozens of genes involved in neuronal communication. That fits the clinical observation that cerebellar abnormalities in tuberous sclerosis are strongly linked to TSC2 mutations, and it narrows the question to a specific population of neurons.

The second is Dctn1, a gene involved in moving cargo inside neurons; its human counterpart DCTN1 is mutated in Perry syndrome. Disrupting it raised expression of cholesterol-synthesis genes. The agents linked that to literature suggesting faulty cholesterol transport drives more synthesis, giving a testable hypothesis about cholesterol distribution in neurons and neurodegeneration. Experts rated it novel and worth following up.

The single-agent ablation is the real result

The most useful number in the post comes from the control experiment. Holding the model and resources fixed, they gave one agent the whole dataset and compared it with the distributed setup. As the workload grew, the single agent searched less and less of the literature: at the largest workload it cited 11 references, while the distributed agents cited more than 700. In blinded comparisons, experts preferred the distributed reports 75% of the time.

The post does not explain why the single agent searched less, so I will not guess at the mechanism. It does match what I see in production: an agent with a very large scope triages rather than going deeper. Narrow units of work (here, one perturbation) give each item dedicated search and reasoning, as long as compute grows with the number of units. It is a map-reduce over hypotheses, where the reduce step is a structured artefact rather than a summary of summaries.

What I would want to know

How many agent reports were confidently wrong, and how did the pipeline catch them? How do the citations hold up under verification, not just in count? Agreement with blinded experts on novelty is encouraging, but it is not experimental validation, and the post does not report any of these leads being tested in the lab yet. The authors are careful here: the output is a list of experiments worth running, and scientists decide which to pursue.

Why it matters

Perturbation screens, cell atlases and large-scale omics now produce far more structure than any lab can interpret. The bottleneck has shifted from producing data to reading it against decades of scattered literature. This work shows a credible pattern for that problem: break the data into units that each fit within one agent's reasoning budget, give every agent the same tools and procedure, and merge the results into something a human can query. I expect it to carry to any domain with many independent hypotheses and a large body of prior work.