What Happened
Anthropic announced on September 23 that its Claude agents had turned up a previously uncharacterized enzyme system inside bacterial-virus genomes. The company calls it array-associated reverse transcriptases, or ART. The system has three parts that travel together: a reverse transcriptase enzyme, a partner gene, and an array of evenly spaced DNA repeats that looks, at a glance, a lot like the repeat arrays found in CRISPR.
The numbers behind the search are the striking part. Roughly 950 agents worked for 21 hours and used about 210 million tokens while scanning more than 200,000 reverse transcriptases drawn from public sequence data. That pass narrowed the field to around 3,500 candidate systems, and the agents condensed their strongest leads into 20 detailed reports for human scientists to read. Independent coverage of the search reports the same headline figures.
People handled everything that touched a bench. Anthropic says its scientists ran the follow-up experiments themselves at biosafety levels 1 and 2, and that the initial experiments are finished while more work is under way. The organisms of interest are bacteriophages, the viruses that infect bacteria, where attack and defense systems are often packed into compact genetic neighborhoods.
What the team has not shown is what ART actually does. The announcement is candid that the primary function is still unknown and that the findings are early. The result is best read as a map of a new place worth exploring, not a finished expedition.
Why It Matters
Reverse transcriptases copy RNA into DNA, and they show up in some of biology's most consequential machinery, from retroviruses to the mobile genetic elements that genome-editing researchers study. A repeat array sitting next to one is exactly the kind of pattern that makes molecular biologists lean in, because the last time repeats and enzymes were found living together, the pairing grew into a technology that reshaped medicine and agriculture.
The second reason is the shape of the method. Genome mining is an old discipline: researchers comb sequence databases for unusual gene neighborhoods and then test the strangest ones. What changes here is the labor model. A swarm of agents can read, cluster, compare and write up thousands of candidates in a day, while a small human team decides which few deserve bench time. If that pattern holds, the bottleneck in discovery moves from finding leads to validating them, and validation is slow, expensive and physical.
The third reason is the claim itself. MIT Technology Review's piece on the episode asks when it is fair to say an AI made a scientific discovery at all, noting that AI companies increasingly describe their systems as discoverers rather than tools. That framing matters because it shapes funding, credit and expectations. It also makes real progress harder to recognize when every interesting pattern is presented as a breakthrough.
Reaction
Reaction split along a familiar line. Lucas Harrington, who trained under Nobel laureate Jennifer Doudna and co-founded Mammoth Biosciences, argued that scanning genomes for linked genes and repeats is routine work and that comparable systems have been known since 2008, according to The Decoder's report. His core point was that spotting a cluster is the easy step, while working out what it does is the hard one.
Feng Zhang, the MIT and Broad Institute scientist best known for pioneering CRISPR genome editing, took a more generous view after reviewing the findings. Anthropic's write-up says he found the results intriguing and sees them as an example of how AI agents can contribute to biological discovery.
The MIT Technology Review article also draws on Mario Rodríguez Mestre, a biologist at the University of Copenhagen, as it examines where the line between discovery and routine lab work sits. The debate is a useful reminder that specialists can look at the same result and weigh it differently depending on what they count as the finish line.
Both camps are describing the same facts. Nobody disputes the 950 agents or the 200,000 sequences. The disagreement is whether finding a pattern deserves the word discovery before anyone knows its function.
What's Next
The obvious next step is function. Until experiments show what the reverse transcriptase and its repeat array do together, whether they defend a phage, edit DNA, record information or do something stranger, ART remains a catalog entry. Anthropic says further characterization is underway.
Phages are a forgiving place to look. They are abundant, their genomes are small, and the arms race with bacteria has produced a long list of molecular tools, including restriction enzymes and CRISPR itself. A team testing ART would likely start by expressing the system in a model bacterium and watching how phage infection changes.
Watch also for outside replication. If the sequences and reports are shared, independent labs with their own bench time could pick up some of the 3,500 candidates and either confirm that ART is distinctive or show that it fits into known families. Either outcome would settle more of the argument than another round of commentary.
Closing Thoughts
The most useful way to hold this story is to separate three questions that tend to get blurred together: did the method work, is the finding new, and does it matter. On the first, the answer looks like yes, since the agents produced a short list that experienced biologists considered worth a look. On the second, experts disagree about novelty. On the third, nobody knows yet, because the function is still unknown.
There is also a quieter lesson about language. As AI labs move from chatbots to research programs, words like discovery will be tested against a slower clock made of peer review, replication and use. The fairest scorecard is one written a year from now, when the lab results are in. Until then, a fast search that surfaces a strange enzyme beside a strange repeat array is a good reason to keep looking and a poor reason to declare victory.
한글 요약
Anthropic은 9월 23일, 약 950개의 Claude 에이전트가 21시간 동안 20만 개가 넘는 역전사효소를 훑어 박테리오파지에서 CRISPR와 닮은 반복 서열을 가진 새로운 효소 시스템(ART)을 찾았다고 발표했습니다. 후보는 약 3,500개로 좁혀졌고, 에이전트가 정리한 20편의 보고서를 연구자들이 검토했으며 실험은 모두 사람이 진행했습니다.
반응은 엇갈렸습니다. 재래식 유전체 탐색에 불과하다는 비판이 나왔고, 비슷한 시스템은 2008년부터 알려져 있었다는 지적도 있었습니다. 반면 장펑은 흥미로운 결과라고 평가했습니다. 핵심 쟁점은 기능이 밝혀지기 전에 이를 발견이라고 불러도 되는가입니다.
ART의 실제 기능은 아직 알려지지 않았고 추가 특성 분석이 진행 중입니다. 독립 연구실의 재현 실험과 기능 규명이 이루어져야 이번 성과의 무게를 제대로 가늠할 수 있습니다. 참고: Anthropic 발표, The Decoder, MIT Technology Review.