Claude Agents Find a CRISPR-Like Enzyme System in 21 Hours

AI agents have just helped find something new in biology. On September 23, 2026, Anthropic said a swarm of Claude agents searched through hundreds of thousands of viral and bacterial enzymes and turned up a previously undescribed genetic system with a family resemblance to CRISPR. Human scientists in Anthropic's own wet lab then ran the first experiment on it, and the early result lines up with what the agents predicted.

• Scale of the search: roughly 950 Claude agents ran for 21 hours and used about 210 million tokens
• The funnel: more than 200,000 reverse transcriptases screened, 3,500 new candidate systems, 20 written up for scientists, one standout discovery
• The discovery: array-associated reverse transcriptase (ART), found mostly in bacteriophages
• First lab result: the ART array is expressed as a set of distinct short RNAs, a CRISPR-like trait

What Is the ART Enzyme System?

ART has three parts sitting side by side in the genome: a reverse transcriptase enzyme, a neighboring partner gene whose job is still unknown, and a long run of evenly spaced DNA repeats. That repeat pattern is what caught attention, because it looks a lot like a CRISPR array. In CRISPR, those arrays are transcribed into short guide RNAs, and those guides are what make CRISPR programmable.

That is why the first wet-lab result matters. When Anthropic's scientists expressed the system, the array came out as a set of distinct short RNAs, just as a CRISPR array does. It is an early and encouraging sign, not a finished story. Anthropic is clear that it does not yet know what ART does in nature, and says the system could potentially support new ways to cut, copy or paste DNA. That is a hypothesis to test, not a capability anyone has shown.

How 950 Claude Agents Split the Work

The interesting part for anyone following AI research news is the division of labor. Claude agents handled the literature review, the pattern-finding across a huge pool of protein sequences, and the hypotheses about which candidates deserved a closer look. They boiled 3,500 candidates down to 20 readable reports, the kind of shortlist a human team can actually act on.

People then took over at the bench. Anthropic set up a Bay Area wet lab in spring 2026, limited to lower-risk BSL-1 and BSL-2 work, and its scientists expressed and tested the protein. According to TechCrunch, Anthropic CEO Dario Amodei noted that Claude does not run the experiments itself today. The search, in other words, was automated; the verification was human.

What the Discovery Does and Doesn't Claim

Good science reporting means keeping the caveats attached. Parts of this system had been glimpsed before: the reverse transcriptase itself appeared in earlier studies of a jumbo phage, and TechCrunch notes that Stanford researchers had separately described a similar system. What the agents appear to have added is noticing the array and the partner protein together, and flagging the combination as worth testing. CRISPR pioneer Feng Zhang of MIT and the Broad Institute reviewed the findings and called the work an exciting example of AI agents contributing to biological discovery. Anthropic has published a technical preprint for other labs to examine.

Why AI-Driven Biology Discovery Matters

The practical lesson is about search. Modern sequencing has produced far more protein data than people can read, and agentic AI turns that backlog into a ranked list. A 21-hour run that ends with a lab-ready shortlist is a genuinely different pace of work for biology.

It also fits a pattern we have been tracking. Anthropic recently opened its Life Sciences Verification program for research labs, and frontier models such as Claude Opus 5.5 keep getting better at long, multi-step agent work. ART is the first time that combination has produced a headline biology result, and the preprint means other scientists can now check the work and build on it.

Sources: Anthropic — September 23, 2026; TechCrunch — September 23, 2026; Al Jazeera — September 24, 2026.