Anthropic researchers announced on September 23 that their AI model, Claude, autonomously identified a previously uncharacterized enzyme system in bacteriophages. Dubbed “ART” for array-associated reverse transcriptases, the system features a DNA repeat architecture reminiscent of CRISPR, though its biological function and potential as a gene-editing tool remain unproven.
How the AI Agent Swarm Identified ART
The discovery emerged from a high-level prompt to search public genomic databases for new reverse transcriptases, enzymes that convert RNA into DNA. Rather than relying on a single process, Anthropic deployed a swarm of approximately 950 specialized Claude agents to process the data. This effort was conducted within a new life sciences research group formed in the spring of 2026 to evaluate whether general AI models could systematize and accelerate discovery.
Over the course of 21.5 hours, the agents analyzed more than 200,000 known reverse transcriptases, utilizing 210 million tokens of data. The AI narrowed this vast pool to 3,500 candidates, eventually flagging 20 as particularly promising for human review. The team noted that what Claude appeared to identify was the significance of the surrounding arrangement, specifically the repeat array and an additional accessory protein whose function remains unknown. This full process, which would typically require human scientists to spend months on manual analysis, was completed by the agents in under 22 hours.
Scientists Synthesize System Amid Peer Review Skepticism
Following the computational identification, scientists at Anthropic’s molecular biology laboratory in the San Francisco Bay Area synthesized the system to test Claude’s hypothesis. While the company has released a technical research paper, the work has not yet undergone peer review.
The announcement has drawn mixed reactions. Kevin Esvelt, Associate Professor of Media Arts and Sciences at the MIT Media Lab and leader of its Sculpting Evolution group, was more skeptical, noting, It looks like a standard CRISPR-like bioinformatics ID. Not sure what the fuss is about.
“Let’s say we are on Santa Monica Beach and trying to scan through all the sand to find a diamond. AI found this thing that looks very shiny, and then you have to go back to the lab to know—is it glass?”
Kevin Blake, a microbiologist at Washington University School of Medicine, observed that naturally occurring CRISPR systems differ sharply from the technologies scientists later developed from them. Researchers have cautioned against assuming ART will become a comparable technology. Michal Rosen-Zvi, the former Director of AI for Healthcare and Life Sciences at IBM Research, has suggested that biology may be next in line for an AI-driven shift, noting that the industry is only at the beginning of that curve.
Questions Surrounding Data Usage and Credit
The discovery also prompted discussion regarding the origins of the data. Mario Rodríguez Mestre, a PhD candidate in computational biology at the University of Copenhagen, said his team has been investigating these enzymes, which they call “jumbotrons,” since 2022. According to The Scientist magazine, Mestre shared his dissertation and manuscript drafts with Claude. Peter Yoon, who previously worked in the lab of CRISPR architect and Nobel laureate Jennifer Doudna, noted that Anthropic’s agents were guided by four senior molecular biologists and domain experts out of a six-author team.
Anthropic Seeks Partners as IPO Nears
Anthropic has not demonstrated that ART can cut, copy, or paste DNA. The company is currently inviting academic and enterprise partners to work on determining ART’s exact primary function. The discovery arrived as Anthropic prepares for an IPO, having submitted a draft S-1 to the SEC on June 1 with a reported November listing target.
The announcement triggered market volatility. Shares of gene-editing companies traded lower on the day of the announcement, though some names recovered part of the loss the next day as it became clear Anthropic had not shown ART can edit DNA. The company maintains that because Claude generates hypotheses so prolifically, the researchers learn which ones are worth testing, and that judgment is fed back into the model’s instructions to improve its scientific utility.