Anthropic announced on September 23 that a batch of Claude agents had found a previously undescribed class of enzyme systems while combing through public DNA sequence databases, naming it ART (Array-Associated Reverse Transcriptases). These systems turn up mainly in bacteriophage genomes, and their structure bears some resemblance to CRISPR. The company also released a preprint for peer review.
What's confirmed so far is the structure, not the function. ART's primary role remains unknown; Anthropic's own wording is that the relevant research is still "ongoing."
950 agents, 21 hours
According to the process Anthropic disclosed, researchers supplied only an initial prompt, leaving the database searches, candidate screening and hypothesis generation to Claude agents.
After 21 hours spent searching this data by roughly 950 agents using 210 million tokens, one of the agents spotted something remarkable.
In plain terms: about 950 agents spent 21 hours searching the data with 210 million tokens before one of them spotted something unusual.
The funnel narrowed like this: the agents first collected more than 200,000 reverse transcriptase sequences, picked out roughly 3,500 new candidate systems, then compressed that down to the 20 most compelling candidates for a report handed to humans.
The selected systems consist of three parts: a reverse transcriptase, an adjacent companion gene, and a long stretch of evenly spaced DNA repeats. Anthropic says Claude appears to be the first to notice both defining features together - the noncoding array of DNA repeats and the extra accessory protein.
Early experiments show ART arrays are expressed into a set of distinct short RNAs, a behavior similar to how CRISPR arrays work - but the company hasn't drawn any conclusions about what these short RNAs actually do inside the cell.
Every experiment was run by humans
The wet-lab portion of this discovery took place in an ordinary molecular biology lab in the Bay Area, with every experimental step carried out by human scientists; Claude handled only the upstream searching, hypothesis generation and data interpretation. The lab works exclusively with BSL-1 and BSL-2 samples and doesn't handle human pathogens.
That's a separate story from the wet lab reported earlier this month, which described Claude directing robots to run experiments. There's no automated execution step in what was disclosed this time - the humans stood at the bench, the model stayed in the database.
Feng Zhang, one of the founders of CRISPR gene editing (MIT and the Broad Institute), called the work:
"This is an exciting example of how AI agents can contribute to biological discovery."
Set against 'AI-designed phages'
Last September, the Arc Institute and a Stanford team used the genomic language model Evo to generate a batch of phage genomes, 16 of which were able to infect and lyse E. coli in lab tests. That work was about "writing" - getting a model to generate sequences that don't exist in nature, then testing in the lab whether they're viable.
ART takes the opposite path - "reading." The sequences were already sitting in public databases; what was missing was someone to go through 200,000 reverse transcriptases one by one and notice that a small subset kept showing up next to a repeat array. This time the job went to a machine, and it took 21 hours.
The two approaches have different bottlenecks. For generative design, the bottleneck is experimental validation - most sequences a model writes don't turn out to be viable. For database mining, the bottleneck is deciding which leads to chase first; 950 agents ran in parallel, but it still came down to 20 reports, with humans deciding which experiments to actually run.
CRISPR's own history offers a reference point: a Japanese team first noticed that odd stretch of repeated sequence in the E. coli genome back in 1987, and it took more than two decades before it was turned into a gene-editing tool. ART is currently at a similar "noticing the repeats" stage - whether it develops into an application depends on functional studies and independent replication.
The preprint is now public. Whether other labs can independently replicate ART's short RNA expression is the first test of this result from the outside.
Sources: Anthropic's official blog, CocoLoop, the ART preprint; agent counts, token usage and funnel figures per Anthropic's disclosure, Evo phage-design figures per the Arc Institute's published paper.