Playbooks · Coding tasks
Thousands of agents read a mouse-brain atlas for leads
Written November 4, 2026
Thousands of agents, each reading one gene disruption, interpreted a whole-brain mouse atlas in 32 minutes and proposed new research leads. The result is a reading of existing data, not a treatment, and the animals were mice.
What changed
OpenAI, PerturbAI, and Scripps Research ran thousands of GPT-Rosalind agents over an existing whole-brain CRISPR atlas — experimental data from millions of mouse brain cells — and combined the agents' reports into what they call a reasoning atlas. Each agent took one gene disruption, examined the cellular responses, searched the literature, and proposed an explanation. The whole interpretation took 32 minutes. OpenAI's account compares that with an estimate that human analysts, working one perturbation at a time, would need months to a year for the same reading. The work is reported in a preprint on bioRxiv, not a peer-reviewed journal article. Two of its authors are co-founders of PerturbAI, and several authors are OpenAI and PerturbAI employees — a conflict disclosed in the preprint.
What the swarm found
- A difference between two similar genes. Disrupting Tsc1 and Tsc2 had similar effects in several neuron groups, but in the cerebellum, disrupting Tsc2 altered dozens of genes involved in neuronal communication, while disrupting Tsc1 changed little. The agents flagged a region-specific difference a single-gene reading could have missed.
- A cholesterol lead for Perry syndrome. Disrupting Dctn1, the gene behind Perry syndrome, increased the expression of cholesterol-making genes. The agents' proposal is a question for follow-up experiments: whether restoring cholesterol distribution could protect neurons. It is a hypothesis, not a finding about treatment, and not a result in people.
- A scale difference against a single agent. With the same model and resources, a single agent cited 11 references at the largest workload, while the distributed swarm cited more than 700. Blinded experts preferred the distributed reports 75% of the time. The comparison shows breadth: thousands of parallel readings cover more ground than one sequential reading.
Strengths
- Parallel reading of an atlas no one person can hold. Millions of cells and thousands of perturbations are beyond sequential human reading. Splitting the work across agents matches the shape of the data.
- Experts judged the output blind. The preference test was blinded: experts did not know which reports came from the swarm. That is a stronger check than the authors rating their own output.
- Leads, stated as leads. The framing throughout is prioritization: which experiments to run next. The DCTN1 proposal is phrased as a question to test, and OpenAI's account says experts considered the hypothesis novel, consequential, and worth following up — the right standard for a lead.
Weaknesses and limits
- Mouse data, not patients. Every cell in the atlas is a mouse brain cell. A lead about cholesterol in mouse neurons says nothing directly about Perry syndrome in people, and nothing about any treatment.
- A preprint with company authors. The work has not passed peer review, and the team that built the system is also the team reporting its success. The blinded expert check helps, but it does not replace independent replication.
- An estimate, not a race. The months-to-a-year human figure is an estimate of what sequential analysis would take. The swarm was not timed against a human team doing the same job.
- Reading data is not generating data. The atlas already existed. The agents interpreted measurements; they ran no experiments, collected no new cells, and proved no mechanism. The DCTN1 idea stays untested until a lab tests it.
- No cost is stated. OpenAI's account gives no figure for the compute behind thousands of agents running for 32 minutes. Anyone planning a similar run cannot budget from this post.
What to do with it
For research teams sitting on large perturbation datasets, the usable pattern is the split: one agent per perturbation, literature search included, reports combined into a shared atlas, experts judging the output blind. That pattern is worth copying as an experiment design, with the same blind check, before trusting it as a pipeline. For everyone else, the right response is patience: watch whether the DCTN1 hypothesis survives a lab test, and whether an independent team reproduces the swarm's reading. A lead that fails in the lab is still a useful result — it is the experiment, not the agent report, that decides.
Limits of this reading
- The evidence is one preprint and one company post describing the same project. No independent replication exists.
- The expert-preference figure (75%) describes report quality as judged blind. It is not a measure of how many leads were correct.
- Whether the swarm's other leads hold up, and how many were false, is not reported. The published account highlights the successes.
Related playbooks
- The 83% AI diagnosis headline, read properly
- Fingerprinting AI coding agents from 33,580 PRs
- Check where your coding agents store API keys on your own machine
- Claude agents now dream: memory that consolidates
Sources
- https://openai.com/index/scaling-ai-research-for-complex-biological-data
- https://www.biorxiv.org/content/10.64898/2026.10.15.759436v1
Back to: Coding playbooks hub · Also: 3,800 hours of independent speech through a brain implant · The 83% AI diagnosis headline, read properly