Anthropic announced two things at once yesterday: a new life sciences research group with its own lab, and that group’s first result. Claude found a previously undescribed enzyme system in the DNA of bacteriophages. Anthropic calls it ART, short for array-associated reverse transcriptases.
The lab itself was no longer a secret. Reuters uncovered the wet lab a week ago, though the company wouldn’t say what was being worked on there. Now it’s on Anthropic’s own blog.
950 agents, 21 hours, 210 million tokens
Claude got one prompt: search a very large database of DNA sequences for interesting new reverse transcriptases. These enzymes, RTs for short, copy RNA back into DNA. In bacteria, many of them are part of the immune system.
After that, Anthropic says, its people only did the lab work. Around 950 agents combed the data for 21 hours and burned 210 million tokens doing it. They gathered more than 200,000 RTs, filtered out 3,500 candidate systems, and narrowed those down to 20, writing human-readable reports on each. A trained scientist needs weeks to months for that kind of analysis.
Reading the raw sequence next to an odd-looking RT, one agent wrote: “[The DNA next to the RT] is spectacular: I can see by eye a tandem repeat array … that’s a CRISPR-like … repeat array?!” Then it counted the repeats, measured their spacing, compared the layout against known RT systems, searched the literature for any earlier report of the pattern, and filed a report for human review.
What ART is, and what nobody knows
ART has three parts: the reverse transcriptase, a partner gene right beside it, and a long array of evenly spaced DNA repeats. That layout resembles a CRISPR array, the bank of RNA sequences that makes CRISPR-Cas systems programmable. First lab experiments show the ART array is also transcribed as a set of short RNAs.
What the system does, Anthropic doesn’t know, and says so plainly. All that’s clear is this: the combination of features ART carries has only ever turned up in a handful of other systems, and all of those are programmable. They cut, copy or paste DNA.
Feng Zhang of MIT and the Broad Institute, one of the pioneers of CRISPR genome editing, read the pre-print. He calls the finding “genuinely intriguing” and worth further investigation. The underlying RT, from a jumbo phage, had already been identified. What’s new is that somebody noticed the two components sitting next to it.
The lab runs on the same tools you have
The lab is in the Bay Area and, per Anthropic, looks like an ordinary molecular biology lab. Work stays at biosafety levels BSL-1 and BSL-2, and nobody there handles pathogens that can infect humans. All bench work is done by people.
One detail matters if you work with Claude yourself: the researchers use Claude Science and Claude Code, the same tools anyone gets. On top of that sits a harness of their own that coordinates many Claude sessions in parallel. And because Claude produces hypotheses in bulk, the question of which ones deserve a test has itself become an object of study. What the team learns there goes back into the instructions Claude gets.
The hard part was throwing things away
A machine spotting something odd among 200,000 sequences is pattern matching at scale. The demanding work sits on either side of it: turning 3,500 candidates into 20. Anthropic writes that most proposals are eliminated during follow-up review, and that a campaign can end with a single candidate worth testing, or with none. That kind of judgement is exactly what models haven’t been trusted with.
Nobody knows what ART does, all the same. The find is a hypothesis with early lab evidence, not a result. Anthropic says as much, which is the honest version. Whether AI agents really speed up biology gets decided by what becomes of ART, not by yesterday’s announcement.