A Claude Agent Spotted a CRISPR-Like Array in Phage DNA

Anthropic said on 23 September 2026 that a swarm of Claude agents searching a database of DNA sequences identified a previously uncharacterised enzyme system in bacteriophages, the viruses that infect bacteria. Roughly 950 agents ran for 21 hours and consumed 210 million tokens before one of them flagged an unusual repeating pattern of DNA next to the gene for an odd-looking reverse transcriptase, the class of enzyme that copies RNA into DNA. The company's scientists named the system array-associated reverse transcriptases, or ART.

The arithmetic of the search is the substance of the claim. According to Anthropic's account, the agents gathered more than 200,000 reverse transcriptases, picked out 3,500 candidate systems and narrowed those to the 20 most compelling, writing a human-readable report on each with the evidence for a proposed function. Anthropic says an expert scientist doing that genome mining by hand would need weeks to months. Human involvement, the company says, was limited to the initial prompt and the laboratory work.

Why a repeat array matters

ART has three parts: the reverse transcriptase, a partner gene of unknown function beside it, and a long array of evenly spaced DNA repeats. It is the third that drew the CRISPR comparison. In CRISPR-Cas systems, a similar array holds a bank of different RNA sequences, and it is that bank which makes the systems programmable. Anthropic reports that its first experiments show the ART array is also expressed as a set of distinct short RNAs, which it describes as suggesting something analogous may be at play. Anthropic also notes that the underlying reverse transcriptase, found in a jumbo phage, had been identified in earlier studies, and that Claude appears to be the first to notice the array and the accessory protein sitting alongside it.

What the lab is and is not

The laboratory is in the Bay Area and Anthropic says it looks like a typical molecular biology facility. It operates only at biosafety levels 1 and 2 and handles no pathogens that can infect humans. All bench work is performed by human scientists; the company says it has experimented with automating lab work but found the approach poorly suited to the improvised workflows of this kind of research. Reuters reported on 18 September, from unnamed sources, that Anthropic's eventual aim is a setup in which Claude directs robotic lab units. Eric Kauderer-Abrams, the company's head of life sciences, told Reuters that the final test in biology is still real lab work, and that Anthropic is not competing with companies that bring drugs to market.

One detail gives a sense of how the work actually proceeded. Anthropic quotes the agent's own words on encountering the sequence: "[The DNA next to the RT] is spectacular: I can see by eye a tandem repeat array … that's a CRISPR-like … repeat array?!" The agent then counted the repeats, measured their spacing, compared the layout against known reverse transcriptase systems and searched the literature for prior reports before filing for human review.

The pattern is familiar from the physics work Parallax Nexus covered earlier this month, in which algorithms proposed hardware layouts rather than tuning parameters. What is new here is the position in the pipeline: the model chose what was worth looking at, and the humans ran the bench. The honest judgment is that this is a strong result about search and triage, and not yet a result about biology. Nobody knows what ART does. The primary function is under investigation, the technical report is a preprint that has not been peer reviewed, and Zhang's endorsement says the finding merits further investigation, which is a long way from saying it works.

Anthropic acknowledges the timing. It says it published early both to demonstrate Claude's capabilities and to show the research community what it is working on, an unusual thing to admit in a discovery announcement and a reminder that the company is preparing to go public.