OpenAI, PerturbAI, and Scripps: What 1,946 Parallel Research Agents Show

What was announced

On October 21, 2026, OpenAI, PerturbAI, and Scripps Research described a run of 1,946 GPT-Rosalind agents, OpenAI's life-sciences model, over an existing mouse brain dataset. R&D World and Unite.AI covered it the same day. The dataset is PerturbAI's whole-brain in vivo CRISPR Perturb-seq atlas, which Unite.AI describes as about 7.7 million cells covering nearly 2,000 disease-linked genes. Xin Jin of Scripps Research, a PerturbAI co-founder, is the corresponding author.

The companion bioRxiv preprint, "Distributed AI agents enable comprehensive discovery of biological insights at scale," was posted October 16 and has not been peer reviewed. Code is on GitHub under jinlabneurogenomics/distributed_agent.

The architecture: one agent per question

The authors call the method distributed biological reasoning. The structure is a fan-out and merge.

1. Split. Each gene disruption becomes an independent task.
2. Fan out. One GPT-Rosalind agent per task examines how cells responded, searches the literature, and proposes an explanation.
3. Merge. The system combines the reports into a reasoning atlas linking genes, cellular effects, papers, and candidate mechanisms, then ranks leads for follow-up.

All agents ran in parallel and finished in 32 minutes. The authors estimate the equivalent human work at months to a year, but R&D World notes that figure is built from assumed times (15 minutes per disruption, 10 minutes per paper), not a measured comparison.

Why fan-out beat a single agent

The baseline was a single agent with the same model, prompt, tools, and data access. As the workload grew, its depth collapsed. R&D World reports that at the largest workload the single agent averaged about 11 references, while the distributed agents retained hundreds of times more. Unite.AI reports that in blinded comparisons, human experts preferred the distributed reports 75 percent of the time.

The mechanism is not mysterious. One agent asked to cover nearly two thousand perturbations spreads a fixed context and attention budget thinly, summarises early, and stops reading papers. Isolated agents each get a full budget for one narrow question. The cost moves to the merge step, which has to reconcile overlapping or contradictory reports.

The pattern generalises to any wide, shallow-per-item workload: auditing every service in a codebase, reviewing every contract in a data room, triaging every alert. If the items are independent, one agent per item with a disciplined merge beats one heroic agent.

How the output was checked

Two checks are worth copying.

Recovery of known results. R&D World reports the agents matched or refined 48 of the atlas study's 55 findings that could be drawn from the data they received, without seeing the original manuscript.

Blinded expert review. R&D World reports 30 doctoral-level biologists reviewed outputs blind, and their judgements broadly supported the agents' rankings of leads.

The highlighted lead: disrupting Dctn1, a gene whose mutations cause Perry syndrome, raised expression of cholesterol-synthesis genes. The agents proposed a cholesterol transport problem as a possible cause. That is a testable hypothesis, not a finding.

Caveats to weigh

• Not peer reviewed. The preprint has not been through review.
• Conflicts of interest. Unite.AI reports six authors are OpenAI employees and two are PerturbAI employees, and that PerturbAI owns the atlas. The preprint discloses this.
• Estimated, not measured, speedup. The months-to-a-year comparison rests on assumed task times.
• Restricted model. OpenAI says GPT-Rosalind access is limited to approved users.

The authors themselves frame scientists as making the decisive calls while agents broaden the questions researchers can pursue. That is the honest reading.

The reusable part

The most durable output is not the 32-minute run; it is the atlas. Once nearly two thousand literature-grounded analyses exist as a structured, linked resource, later questions can retrieve them instead of rerunning the agents. Expensive reasoning, stored and indexed, becomes cheap lookup. Any team running large agent fan-outs should plan where the results will live and how the next agent will find them.