950 Claude Agents Run 21 Hours to Discover Novel Enzyme System ART with CRISPR-like Repeats

What Was Discovered

The system is named ART (Array-associated Reverse Transcriptase) . It appears primarily in phages and consists of three parts: a reverse transcriptase, an adjacent companion protein gene, and a long array of evenly spaced DNA repeat sequences. Phages are viruses that infect bacteria; reverse transcriptases copy RNA information into DNA.

The key insight was not merely finding a previously unseen enzyme gene — related reverse transcriptases were already known. Claude noticed the repeat sequences and companion protein next to the reverse transcriptase and linked all three as a potentially unrecognized system.

The repeat arrangement resembles CRISPR. The team subsequently detected that it produces multiple distinct short RNAs — a step beyond DNA arrangement alone — but what these short RNAs actually do remains unknown and requires further experiments.

How Claude Found It

Anthropic tasked Claude with searching large-scale DNA data for reverse transcriptases worth studying. The search employed approximately 950 agents running for 21 hours , consuming 210 million tokens . They aggregated over 200,000 reverse transcriptases, filtered ~3,500 candidate systems, and produced reports on the top 20 for human researchers.

The preprint states the original search used Claude Mythos 5 . One agent directly examining DNA sequences noticed a reverse transcriptase gene with regular repeat fragments nearby. It then counted repeats, measured spacing, compared against known systems, and consulted literature before handing the finding to human researchers for verification.

In this workflow, Claude performed broad search, proposed leads, and organized evidence. Laboratory validation was done by scientists . Anthropic built its own molecular biology lab rather than relying solely on model reading of papers to claim discovery.

How Far from "AI Invents Gene Editing Tool"?

ART cannot yet be called a new gene editing tool. It has CRISPR-like repeats, and other systems with similar features have shown programmable capability; however, whether ART can specifically recognize, cut, or rewrite DNA has not been proven . Anthropic explicitly states they still do not know its primary biological function.

The preprint reveals a telling detail: the team ran the same search pipeline 10 more times. Most reruns encountered the ART-related region but did not read the upstream DNA containing the repeats, so they failed to recognize it again. This discovery path is currently not stable.

Claude has already helped researchers pick a promising experimental target from massive gene data. Whether ART can become a useful biotechnology tool awaits functional experiments and independent replication.