950 Claude Agents Scoured 200,000 Enzymes in 21 Hours and Shook Wall Street

Almost a thousand AI agents, 1.9 billion protein clusters, 21 hours of compute, and one mysterious DNA pattern nobody had noticed before. That's Anthropic's headline-grabbing claim this week, and it has already rattled gene-editing stocks. But the full picture is far more nuanced than the hype suggests.

How Claude found ART in bacteriophage DNA

Anthropic gave its Claude agents a broad instruction: search a massive DNA sequence database for unusual reverse transcriptases, enzymes that copy RNA into DNA. Roughly 950 agents ran for 21 hours, consuming about 210 million tokens. They screened over 200,000 reverse transcriptases, shortlisted 3,500 candidates, and distilled those down to 20 systems worth a close look.

One agent spotted something striking: a repeating DNA pattern next to an atypical reverse transcriptase gene. The layout mirrors CRISPR arrays, which store RNA guides that make gene editing programmable. Anthropic named the three-part package ART (array-associated reverse transcriptases).

What ART actually does remains a total mystery

Here's the critical caveat. Nobody knows what ART does. Anthropic says so explicitly. CRISPR pioneer Feng Zhang, who reviewed the preprint, called it "genuinely intriguing" but stressed that further investigation is needed. Computer scientist Dimitri Perrin noted that the enzyme itself was already known; what's new is the recognition that it may form part of a larger system.

Reproducibility is also a concern. When Anthropic repeated the search ten more times, none of the runs found the repeat array. The company attributes this to the vast search space and inherent variability in how agents navigate it. That's honest, but it's not exactly reassuring for peer review.

Gene-editing stocks dropped before anyone verified the science

Wall Street moved fast. CRISPR Therapeutics fell about 5.5%, Beam Therapeutics dropped 6%, and Prime Medicine slid 11.6% on the day of the announcement. The sell-off reveals how jittery biotech investors are, even when the discovery hasn't been shown to edit a single gene.

This is the first published result from Anthropic's new Bay Area wet lab, which operates at biosafety levels 1 and 2 with all physical experiments conducted by human scientists. The timing is notable: Anthropic is preparing its IPO, and a splashy biology result doesn't hurt the narrative.

The real story: AI as a scientific search engine

ART itself may or may not lead anywhere. The more lasting takeaway could be the method. A general-purpose language model like Claude sifted through a dataset that would take a PhD student weeks to explore, flagged a candidate, and generated human-readable reports. OpenAI has entered the same arena with its own life-sciences model. The race to make AI a genuine lab partner is on.

Still, let's keep perspective. An unreviewed preprint, a non-reproducible result, and an unknown biological function do not add up to "the next CRISPR." They add up to a promising lead and, perhaps more importantly, proof that AI agents can handle the grunt work of genomic exploration at scale.

What is the ART enzyme system discovered by Claude?

ART (array-associated reverse transcriptases) is an enzyme system found in bacteriophage DNA. It consists of a reverse transcriptase, a partner gene, and evenly spaced DNA repeats resembling CRISPR arrays. Its biological function is still unknown.

Has Anthropic's ART discovery been peer-reviewed?

No. Anthropic published the results as a preprint on September 23, 2026, without peer review. The company also reported that ten subsequent attempts to reproduce the AI-driven discovery all failed to find the same repeat array.