OpenAI says agents proposed brain leads in 32 minutes. Write the test before you follow one.

On October 21, 2026, OpenAI reported that it, PerturbAI, and Scripps Research used thousands of GPT-Rosalind agents to propose explanations for an existing whole-brain CRISPR atlas. OpenAI says the run took 32 minutes and the atlas draws on millions of mouse brain cells, both company-reported. The day before, OpenAI said it had acquired Orchestra Bio. The standing rule is which decisions an AI should never take alone.

This is an illustrative reading of those two posts through the Of One method. It adds no facts, and it recommends no experiment, treatment, or program bet.

What OpenAI reported

Each agent, OpenAI says, took one gene disruption, read the cellular response, searched the literature, and proposed an explanation. OpenAI says the reports form a reasoning atlas of genes, cellular effects, publications, and possible mechanisms, and that comparable human analysis and literature review, one perturbation at a time, would take months to a year, a company-reported estimate. OpenAI says the agents reproduced findings from human analyses, surfaced further leads, rated novelty and scientific value in line with blinded human experts, and that the findings help prioritize experiments that test the mechanisms. OpenAI says other scientists can explore the atlas.

OpenAI says Tsc1 and Tsc2 work together to control mTOR, and that disrupting either gene had similar effects in several groups of neurons. In the cerebellum, OpenAI says, Tsc2 disruption altered the activity of dozens of genes involved in neuronal communication, while Tsc1 disruption produced little change. Dozens is OpenAI’s count. OpenAI says cerebellar abnormalities in tuberous sclerosis are strongly linked to TSC2 mutations, and that the agents pinpointed a specific group of neurons and two lab questions: which changes existing medicines might correct, and which might point to a new treatment approach.

OpenAI says DCTN1 mutations cause Perry syndrome, a neurodegenerative disease with Parkinson-like symptoms, and that the gene helps move cargo inside neurons. In the mouse atlas, disrupting Dctn1 increased expression of genes involved in making cholesterol. OpenAI says the agents connected that rise with research suggesting that problems with cholesterol transport may cause cells to make more cholesterol. OpenAI’s follow-ups are questions: whether distribution within neurons changes, whether that could contribute to neurodegeneration, and whether restoring a normal distribution could help protect neurons. OpenAI says human experts called the hypothesis novel, consequential, and worth following up.

Sarah Teichmann, identified by OpenAI as Chair in Stem Cell Medicine at the University of Cambridge and founding co-chair of the Human Cell Atlas, said: “AI could help more researchers explore these resources and turn unexpected observations into new questions to test.”

What the 75% measures

OpenAI says it held the model and the resources fixed and compared the distributed agents with one agent given the whole workload. At the largest workload the single agent cited 11 references and the distributed agents cited more than 700, both company-reported. In blinded comparisons, human experts preferred the distributed-agent reports 75% of the time, company-reported.

Of One keeps the 75% as evidence about which report experts preferred. The mechanism still needs a kill test: the result that would prove the claim wrong. Papers that arrived through one agent reading are one chain of support. of1.ai’s line for a skipped label is “Every question you skip becomes a guess treated as a fact.”

Who signs the program response

On October 20, OpenAI described Orchestra Bio as a platform where scientists and program leaders delegate monitoring and tracking of studies and campaigns to agents. The agents help teams identify affected activities and timelines, gather the relevant information, and compare ways to respond. A failed manufacturing batch that can delay a clinical trial is the example OpenAI gives. Scientists and program leaders decide what to do next.

Of One puts a named person on moves that touch health, money, or physical safety, and on anything you cannot undo. The gate is the response that spends money, moves a trial, or reaches a patient. The signer should have the observation, the inference, and the kill test. If a gap stays open, accepted risk names the person, the scope, the expiry, and what would reopen the question. The method leaves the check visible while there is time to run it, and it does not promise to catch every error.

A card you can copy

The lines below are a reading of OpenAI’s Dctn1 paragraphs.

• Claim the agents raised. Dctn1 disruption may change how cholesterol is distributed within neurons, and that change may contribute to neurodegeneration.
• Observed, as OpenAI reports it. Dctn1 disruption raised expression of genes involved in making cholesterol.
• Still a question. The transport explanation, a role in neurodegeneration, and a benefit from restoring distribution.
• Kill test. If a follow-up shows no change in how cholesterol is distributed within neurons, the distribution claim stops until some other result supports it.
• Decider. Researchers choose the experiment. OpenAI names scientists and program leaders for the program response.
• Gate. Spending money, changing a trial, or reaching a patient waits for that named person.
• Reopen. A distribution measurement, or a result that cuts the link to neurodegeneration.

What to do this week

• Split one AI write-up that is about to steer a meeting into what a dataset or a paper states and what the model inferred.
• Write the kill test beside the inference. A health, money, or hard-to-undo move waits until that test is written down.
• Name the signer. If a gap stays open, write the scope, the expiry, and what would reopen it.
• Keep the 75% on the report comparison. Use the kill test on the mechanism.

You can run the eight starting questions on a decision of your own with the free tool at of1.ai/ask.

If you want claim cards, kill tests, and human gates for a decision your team already faces, that is a consult: $400 an hour, in 5-hour blocks. You leave with the map and a first move you can undo.