Anthropic's Claude Agents Find a CRISPR-Like Repeat Array in 21 Hours

Anthropic said on Sept. 23 that about 950 of its Claude agents, running for 21 hours against a sequence database, flagged a new enzyme system inside the DNA of bacteriophages, the viruses that infect bacteria ¹. The run spent 210 million tokens ¹. They sorted more than 200,000 reverse transcriptase genes down to 20 candidates worth a closer look ⁵. Expert teams usually spread that work across weeks or months ⁵. One agent stopped on a repeating pattern sitting beside an unusual reverse transcriptase gene, and Anthropic named the result array-associated reverse transcriptases, or ART ¹. Humans wrote the prompt. People ran every experiment at the bench, and the function of the system remains open ².

Why does a preprint about phage DNA belong on a ledger about superintelligence? Because the autonomy component asks whether systems run long, useful work at measured reliability. A 21-hour search that ends in a laboratory candidate is a claim on exactly that. Sizing the step is the whole argument.

What Claude's agents ran, and who checked the result

Genome mining is search under a filter. A pipeline pulls a protein family out of public sequence data, clusters it, and surfaces the members whose neighbourhoods look strange. Here the family was reverse transcriptases, the enzymes that write DNA from RNA. The strange neighbourhood was a tandem repeat array of the kind CRISPR systems use to store fragments of past infections ¹. Anthropic's researchers describe a three-part system and say it can perform operations like cutting, copying and pasting DNA ².

Dario Amodei presented the finding to the U.N. Security Council on Sept. 24 and called it preliminary ⁵. He was precise about authorship in Anthropic's own telling, saying the work was done "mostly, though not entirely, by Claude," and he pointed to a Stanford team that "previously discovered a system that is in some ways similar" ². Read the second clause. Amodei, announcing a first discovery from Anthropic's molecular biology lab, spent a sentence placing it beside prior human work, which is the kind of hedge that usually arrives from a reviewer instead.

Feng Zhang reviewed the result for Anthropic. He is the MIT and Broad Institute researcher who turned CRISPR into a working genome-editing tool in human cells, so his read carries weight on both the biology and the framing. "The identification of RNA-repeat arrays associated with reverse transcriptases is genuinely intriguing and merits further investigation," Zhang said in Anthropic's announcement ¹. He praised the pattern, and he left the mechanism open. "This is an exciting example of how AI agents can contribute to biological discovery," Zhang added, closing on a hope that the work encourages more scientists to try the method ¹. Both sentences point forward. Each one stops short of what ART does.

The case for the move

Speed and parallelism are the evidence. A single agent scanning 200,000 genes is a script. Run 950 of them for 21 hours, keep a shortlist, and hand 20 candidates to human biologists, and the job is one a lab group used to spend a quarter on ⁵. Stanley Qi, an associate professor of bioengineering at Stanford, called the discovery "incredibly exciting" and credited the system's ability to recognize an unusual biological pattern that had been difficult to detect before ³. That is the autonomy claim in its cleanest form: pattern recognition at a scale where human attention runs out.

The run also has a property the ledger rewards. Its output went to a wet lab, where people ran experiments and early results suggested the system may be programmable ⁵. A benchmark score arrives inside a vendor's harness and stops there. Candidates that survive contact with a bench have met a test outside the model.

Amodei drew the boundary himself. "Eventually it may even be possible for Claude itself to safely perform the experiments by autonomously controlling lab equipment, with appropriate safeguards in place, but we aren't doing that today," he said ². The sentence concedes where the autonomy stops. Search ran on its own. Everything downstream of the shortlist stayed with people, at biosafety levels 1 and 2 ².

The case for a smaller step

Lucas Harrington makes the strongest version of the objection. He took his PhD under Jennifer Doudna and co-founded Mammoth Biosciences, so he has spent a decade inside the same search space. Writing on X after the announcement, Harrington said the method has been around for decades, and he dated similar systems to 2008 ⁴. "The hard part is figuring out what a system actually does, and Anthropic hasn't shown that," he said ⁴. He added that "Presenting early results as a major discovery isn't helpful" ⁴. Two claims sit in there, and they pull apart. Harrington grants the search. He rejects the framing around it.

Kevin Blake, a microbiologist at Washington University School of Medicine, put the same caution on the application side. "There's nothing to indicate this is a rival to CRISPR-the-technology, or could be developed into any kind of therapeutic or practical application," Blake said ³. Consider what a CRISPR comparison needs to hold: a guide, a nuclease that cuts where the guide points, and evidence the pair works in a cell. Anthropic's announcement carries repeat arrays and an unknown protein ¹. The cutting partner stays hypothetical.

Publication status caps the move as well. Results live in a preprint, and peer review awaits them ⁵. The ledger's method treats a company announcement of its own work as reported at best, and treats a claim whose central variable is undetermined as inferred. Function is that variable here. Until an outside group says what ART does, the finding is a hypothesis with a good pedigree. Pedigree counts for something on this ledger. It counts for less than a measurement.

Where this sits on the ledger

The reading measures four clauses: unsupervised expert work across occupations, ten gigawatts acting as one machine, measured self-improvement, and third-party proof of execution. This run touches the first clause inside one narrow occupation, computational biology, for 21 hours. Compute and energy stay untouched. Self-improvement stays untouched, since the agents searched phage genomes instead of their own training pipeline. Proof of execution stays weak, because the only record of the search is the company's own account of it. Logs, seeds and the shortlist at each stage would let an outsider replay the run. Anthropic published the story of the run instead.

So the step is small and it is real. Autonomy moves 0.3 points at inferred confidence, which is the band the method reserves for a pattern read from several weaker signals. A published function, confirmed by a group outside Anthropic, would earn a larger move in a later piece. Peer review would earn another.

By the numbers

• 21 hours of continuous agent search against a DNA sequence database ¹
• About 950 Claude agents ran in parallel across that window ¹
• 210 million tokens consumed by the run ¹
• More than 200,000 reverse transcriptase genes examined, narrowed to 20 candidates ⁵
• Weeks or months is the human expert timeline the company compares against ⁵
• Biosafety levels 1 and 2 at the Bay Area laboratory where people ran the experiments ²
• 2008 is the year Harrington dates comparable systems to ⁴

What to watch

An independent laboratory characterizing ART, with a named nuclease partner and a demonstrated edit, would turn this inferred step into a reported or confirmed one. Peer review of the preprint is the second marker, and the reviewers' verdict on the search method matters as much as the verdict on the biology. Watch also for the first run where a model designs the follow-up experiment and a person executes it. That step crosses from search into hypothesis generation. Clearest of all would be a second team reproducing the pipeline on another protein family, then reporting the hours and tokens it cost them. Cost per candidate is the number that turns a demonstration into a method.