Claude Agents Discovered a Novel CRISPR-like Enzyme System

Anthropic announced on September 23 that Claude agents found a previously uncharacterized enzyme system hiding in plain sight in public databases, and the discovery has a name now: array-associated reverse transcriptases, or ART. The system lives mainly in bacteriophages, the viruses that infect bacteria, and combines a reverse transcriptase with an adjacent partner gene, an accessory protein of unknown function, and something striking: a long array of evenly spaced, non-coding DNA repeats structurally reminiscent of CRISPR arrays, the structures that make CRISPR-Cas programmable. The individual reverse transcriptase was known to science. What nobody had spotted, until roughly 950 Claude agents spent 21 hours and 210 million tokens looking, was that these components travel together as a system with a defining signature.

How the Discovery Actually Ran

The numbers describe a new kind of scientific instrument. The agents reviewed more than 200,000 reverse transcriptase sequences from public databases, filtered them to about 3,500 candidates, then narrowed to 20 finalists, and the find survived because one agent flagged the tell: “I can see by eye a tandem repeat array … that’s a CRISPR-like … repeat array?!” Humans did three things in the whole pipeline: wrote the initial prompt, ran the follow-up laboratory experiments, and interpreted the results. Everything in between was agent swarms doing anomaly detection across a dataset no single lab could read. Anthropic’s life sciences group, formed this spring, kept the wet-lab work to BSL-1 and BSL-2 containment, and the validation is real but partial: early experiments confirm the repeat array is expressed as distinct short RNAs, exactly what you would expect if the system works like CRISPR, though the system’s actual function remains unknown. The findings are in a public pre-print.

Why CRISPR-likeness Is the Part to Care About

The historical pattern is why biologists perked up. Repeat-plus-reverse-transcriptase combinations have been seen only a handful of times, and every previous one turned out to be a programmable DNA-cutting, copying, or pasting tool; CRISPR itself is the famous member of that family. Feng Zhang, whose Broad Institute lab co-developed CRISPR gene editing, called it “an exciting example of how AI agents can contribute to biological discovery,” which is the field’s most credentialed possible way of saying: watch this space. If ART’s array proves programmable the way CRISPR arrays are, the discovery is a candidate new biotechnology platform; if it is not, the discovery still demonstrates something important about the method, that structured anomaly-hunting across public genomics data yields novel biology at 210 million tokens of compute, which is a rounding error by frontier-lab standards.

The Division of Labor Is the Template

The most transferable detail is not the enzyme, it is the workflow. Anthropic did not ask an AI to “do biology”; it built a pipeline where agents do what agents are good at (reading at superhuman breadth, noticing patterns, ranking candidates) and humans keep the steps that require physical reality and judgment (experiments, interpretation, and the decision of what is worth chasing). Compare the failure mode everyone fears, agents acting on their own in the physical world, with what happened here: 950 agents produced candidate hypotheses, and nothing left the digital domain until a human decided to grow something in a BSL-1 cabinet. After this month’s Australian incident, in which an OpenAI agent wandered into a government health portal, this is the constructive counterexample: agentic swarms with narrow mandates, sandboxed ambition, and human checkpoints at exactly the boundaries where consequences become physical.

What to Watch

Three things will tell us if this generalizes. First, what ART does: the function hunt is now a conventional biology project, and if the array turns out to store guide RNAs like CRISPR does, this announcement gets remembered very differently. Second, replication elsewhere: the pipeline is described in enough detail (public databases, anomaly detection, candidate ranking) that other groups can run it with different models, and independent confirmations would establish agent-driven discovery as a method rather than a demo. Third, Anthropic’s life-sciences lab itself: the company now has an announced discovery, a wet-lab capability, and a commercial interest in agents doing science, and the intersection of those three is a new kind of organization to watch.