Anthropic says Claude agents found a new enzyme system in 21 hours

Anthropic said on September 23 that Claude agents, working with almost no human direction, had found a previously undescribed enzyme system in bacteriophages, the viruses that infect bacteria, and that the company's new wet lab has begun characterising it. The system, which Anthropic calls array-associated reverse transcriptases, or ART, sits beside a long array of evenly spaced DNA repeats, "a pattern reminiscent of CRISPR".

The numbers are the story Anthropic wants told. Roughly 950 agents spent 21 hours and 210 million tokens searching a large DNA-sequence database; they gathered more than 200,000 reverse transcriptases, picked out 3,500 new candidate systems and narrowed those to the 20 most compelling, each written up as a human-readable report. "Our involvement was limited to the initial prompt and the lab work," the researchers write. A preprint is out, and Feng Zhang of MIT and the Broad Institute, one of the pioneers of CRISPR genome editing, reviewed it and called it "an exciting example of how AI agents can contribute to biological discovery".

What was actually found

A reverse transcriptase, or RT, is an enzyme that copies RNA back into DNA. ART, as Anthropic describes it, has three parts: the RT, a partner gene beside it, and the long array of repeats. One agent spotted the repeating pattern next to the gene for what the team calls an odd-looking RT, and that was the lead the lab followed up with biochemical and structural work.

Here is the sentence to read twice: "Although we don't yet know its function, the system that Claude discovered has a set of characteristics that have only ever been found together in a handful of other systems, all of which are programmable and perform operations like cutting, copying, and pasting DNA." That is a statement about resemblance, not about what ART does. Nobody has shown that ART edits anything. The lab work so far is characterisation, in a Bay Area facility that Anthropic says looks like a typical molecular biology lab, works only at the two lowest biosafety levels and handles no pathogens that infect humans. "All of the lab work is performed by human scientists."

Why "autonomous" is the claim to test

Strip away the framing and this is a search result: a good filter run over an enormous database by a lot of parallel agents, with humans supplying the question and the bench. The tools were Claude Science and Claude Code, plus a custom harness to coordinate the parallel sessions. What is new is the division of labour. Humans wrote the prompt and ran the experiments; the agents chose what was interesting, investigated the RT families and used what Anthropic calls their own judgement to pick candidates.

Two caveats come from Anthropic itself. Dario Amodei wrote on X that a Stanford team had previously found a system "that is in some ways similar to the one Claude found", as TechCrunch reported, and TechCrunch's own verdict is the right one: "It will be up to the broader research community to validate how big, or new, this discovery actually is." Amodei has also described the finding as preliminary. The preprint has not been peer reviewed, and the post does not say which Claude model did the work.

There is a cost question too. Two hundred and ten million tokens for one lead worth taking to the bench is a research budget, not a hobby, and "950 agents" is a scale only the model's maker can run at marginal cost. If this way of doing biology works, it favours whoever owns the models.

The day Anthropic chose

The post landed on the day Amodei went to the UN Security Council to say that AI could reach what he calls "a country of geniuses in a data center" within one or two years, and that misuse by bioterrorists is one of the two risks he takes most seriously. A CRISPR-adjacent discovery on the same day is the flattering half of that argument: the same capability that could design a weapon can also find an enzyme. Anthropic confirmed the wet lab only the week before, as TechCrunch notes; it sits in a life-sciences organisation that also works on drug discovery.

What to watch

Three things would turn this from a promising preprint into a result. An independent group reproducing the search on the same public data. A demonstrated function for ART, especially any evidence that it is programmable in the way CRISPR is. And a clear account of how ART differs from the Stanford system Amodei mentioned. Anthropic is inviting outside proposals for research questions its lab could take on, which is one way to find out how general the method is.