A large scientific dataset can contain more promising questions than a research team has time to pursue. That makes attention a practical constraint: which observations receive a careful reading, which papers get connected, and which anomalies remain unexplored?
On October 21, OpenAI described a collaboration with PerturbAI and Scripps Research that assigned thousands of GPT-Rosalind agents to existing data from millions of mouse brain cells. The team reported completing the analysis in 32 minutes. The work offers a concrete case for examining how agents divide research, and how their output should be judged.
Give each agent a bounded question
The project's public repository describes a pipeline that assigns individual gene perturbations to independent agents. A perturbation is a deliberate disruption used to investigate a gene's effects. Each assignment combines cell-type-specific measurements with literature research, producing both narrative and structured reports.
A subsequent synthesis looks across those reports for relationships and possible follow-up hypotheses. This division matters because a useful investigation needs more than a broad summary. It needs a specific observation, an explanation supported by evidence, and a way to compare that explanation with alternatives.
For a team designing a similar workflow, the practical question is where to draw the boundary around each assignment. A unit should be small enough to investigate thoroughly but retain enough context to make its answer meaningful. Splitting work mechanically can hide relationships that only become visible across assignments.
Read the comparison carefully
The associated bioRxiv preprint was posted on October 16. Its authors report that agents recovered conclusions from the source study without access to its manuscript, predicted held-out perturbation effects, and identified relationships between perturbations. Those are different tests: recovering known findings, making predictions, and proposing connections should not be collapsed into one claim about intelligence.
OpenAI also reports that blinded experts preferred distributed-agent interpretations over single-agent interpretations 75% of the time. At the largest workload, the distributed approach cited more than 700 references, compared with 11 for the single agent.
These are results reported by the research team. Citation volume measures breadth, not whether every source is relevant or correctly interpreted. Expert preference is informative, but it does not certify every proposed mechanism. Readers should ask which evaluation corresponds to the capability they actually need.
Keep a research lead provisional
One example involves Dctn1 disruption and increased expression of genes involved in cholesterol production. The agents connected that observation with literature on cholesterol transport, suggesting further experiments on cholesterol distribution inside neurons and its possible relationship to neurodegeneration.
That is a research question with a testable direction. The report does not establish that changing cholesterol distribution will protect human patients.
A useful review practice is to keep three fields separate: the measured observation, the proposed explanation, and the experiment that could challenge it. An explanation can sound persuasive while several alternatives remain open. Asking what result would count against it is often more productive than asking an agent to make its narrative more confident.
Make the work inspectable
The researchers have released a repository containing the agent runtime, corpus tooling, and notebooks for the manuscript's analyses. It also links to a hosted corpus for inspecting findings, claims, and supporting evidence.
There is an important access distinction. According to the repository, the hosted service does not include the raw dataset files and cannot perform operations requiring direct access to, or recomputation from, those data. The local workflow provides a separate route for dataset-level analysis.
For anyone evaluating the release, browsing a finding and reproducing an analysis are therefore different tasks. A sensible first review would select a small sample of claims, follow their supporting evidence, and document exactly which steps can be checked. Public materials create an opportunity for scrutiny; they do not, by themselves, show that independent replication has happened.
Measure the next decision
The most useful adoption question is what happens after the reports arrive. Does a researcher find a worthwhile experiment sooner? How many suggestions survive expert review? How much work is required to reject weak ones?
A small pilot could track those outcomes alongside runtime and cost, preserving rejected suggestions rather than showcasing only the strongest examples. That would make it easier to see whether parallel agents improve the allocation of human attention.
This study makes the possibility tangible. The next milestone is evidence that the additional leads consistently help researchers choose better experiments, with the uncertainty still visible when the decisions are made.




