Claude agents find a CRISPR-like enzyme system Anthropic calls ART

Anthropic announced a new biology research group and Bay Area lab, and said Claude autonomously discovered array-associated reverse transcriptases (ART) — a novel enzyme system with CRISPR-like DNA repeats in bacteriophages — with only high-level human direction.

HEALTH desk — gene-editing tools like CRISPR started as odd DNA patterns someone noticed; Anthropic is saying an AI agent can notice the next odd pattern at database scale, which pushes biology discovery from months of expert genome-mining toward agent campaigns that humans then verify in a real lab.

Anthropic announced a new life sciences research group and laboratory on 23 Sep 2026. The record is the company post, “Claude discovers a novel enzyme system with CRISPR-like repeats.” The page date is Sep 23, 2026. The page this desk read does not print an hour. The post says the group works on fundamental biology with Claude: exploring datasets of DNA to identify uncharacterized protein families, generating hypotheses at scale, and testing them in the lab. A protein family is a set of related proteins. A hypothesis is a testable guess about what a gene does. The post says the group was formed in the spring of 2026. The lab is in the Bay Area. This desk did not visit it.

What the post says the early result is. Claude discovered a novel enzyme system with properties reminiscent of CRISPR, with only high-level direction from the scientists. They named it array-associated reverse transcriptases, shortened to ART. A reverse transcriptase, shortened to RT, is an enzyme that copies RNA into DNA. RNA is a working copy a cell makes from a gene. DNA is the stored genetic text. CRISPR is a repeating pattern first noticed in bacterial DNA. It became a tool for editing genes. Reminiscent means the layout looks like that pattern. It does not mean ART is CRISPR, and it does not mean ART edits a gene. An enzyme is a protein that does a chemical job.

The search, in the post’s numbers, before the preprint’s different counts. Anthropic gave Claude a prompt to search a massive database of DNA sequences for interesting new examples of RTs. The post says human involvement was limited to that initial prompt and to the lab work. After 21 hours, roughly 950 agents, using 210 million tokens, one agent spotted a repeating pattern of DNA next to the gene for an odd-looking RT. A token is a small chunk of text the model reads. An agent, here, is one Claude session that can take a series of steps, not only answer one question. The pattern sits in bacteriophage DNA. A bacteriophage, or phage, is a virus that infects bacteria. These three figures — 21 hours, roughly 950 agents, 210 million tokens — are the post’s. The preprint below uses 21.5 hours, 949 sessions, and 215.6 million tokens. Do not average them.

What ART is made of, on the post. Three parts: the RT, a partner gene beside it, and a long array of evenly spaced DNA repeats. A partner gene is a second gene that sits next to the first. An array is a run of copies. The post says the underlying RT, found in a jumbo phage, had already been identified in earlier studies. A jumbo phage is a phage with a very large genome. Claude, the post says, appears to be the first to notice the system’s defining features: an associated array of non-coding DNA, and an extra accessory protein whose job is unknown. Non-coding means that stretch of DNA is not a recipe for a protein. The first experiments, the post says, show the array is expressed as a set of distinct short RNAs. Expressed means the cell makes RNA copies of it. The post says a CRISPR array holds a bank of different RNA sequences that make CRISPR systems programmable tools, and that something analogous may be at play here. Analogous is the post’s hedge. It is not a measured function.

What the post says is still unknown, and why they published anyway. The primary function of ART is still being worked out. The post says they do not yet know what it does. It also says this kind of characteristic set has only been found together in a handful of other systems, and that those systems are programmable and do jobs like cutting, copying, and pasting DNA. Beyond CRISPR, the post says, several other such systems are now in development as promising tools. Those sentences are Anthropic’s. A family resemblance is not a product, and it is not a medicine.

The sentence the agent produced, and the census the post prints. While reading the raw DNA next to the RT, the agent said, in the post’s brackets and ellipses: “[The DNA next to the RT] is spectacular: I can see by eye a tandem repeat array … that's a CRISPR-like … repeat array?!” A tandem repeat is the same short DNA word copied again and again, with space between the copies. The post says the agents gathered over 200,000 RTs, picked out 3,500 new candidate systems, and narrowed those to the 20 most compelling candidates, each written up as a human-readable report. It says that kind of analysis can take an expert scientist weeks to months. Those counts are the post’s. The preprint’s census is a different list. Do not treat 3,500 and 20 as the paper’s figures.

The outside quote, as a quote. After reviewing the preprint, Feng Zhang, one of the pioneers of CRISPR genome editing and a professor at MIT and the Broad Institute, said this is an exciting example of how AI agents can contribute to biological discovery. He said the identification of RNA-repeat arrays associated with reverse transcriptases is genuinely intriguing and merits further investigation. He said he hopes the work encourages more scientists to explore how AI can support their research. That is his sentence on Anthropic’s page. A quote is not a second experiment. This desk did not interview him.

The lab’s limits, on the same post. The lab looks like a typical molecular biology lab. The work uses only the lower biosafety levels, BSL-1 and BSL-2. Biosafety level is the containment grade of a lab. The post says they do not handle pathogens that can infect humans. A pathogen is a germ that causes disease. All of the wet-lab work is done by human scientists. Wet lab means the bench, the tubes, and the cells, not the computer search. The team sits inside Anthropic’s life sciences organization, next to groups that work on drug discovery and on training Claude in biology and chemistry. They do the computer work in Claude Science and Claude Code, tools the post says any scientist can use, and sometimes with a harness of their own that runs many Claude sessions in parallel. A harness, here, is the software that coordinates those sessions. These lines are the post’s. This desk did not inspect the lab’s permits.

What the same-day preprint counts, and where it does not match the post. The PDF is titled “Autonomous AI agents discover reverse transcriptases with tandem repeat arrays.” The authors on the first page are Peter H. Yoon, Januka S. Athukoralage, Emmanuel Ameisen, Eric Kauderer-Abrams, Nicholas T. Perry, and Matthew G. Durrant, Anthropic, San Francisco. Correspondence on that page goes to nicholasp@anthropic.com and mdurrant@anthropic.com. The abstract says Claude Code instances surveyed RT loci across 1.9 billion protein clusters. A protein cluster is a group of similar protein sequences. The results section says the agents ran with Claude Mythos 5, recovered about 200,000 RT clusters, and a figure counts 198,290 RT clusters after filtering, searched across 1.94 billion protein clusters. They sampled about 11,000 RT loci and scored 3,564 protein families that kept showing up next to those RTs. Sixteen families passed their screen. A follow-up task promoted a seventeenth. The full campaign was 119 tasks and 949 agent sessions: 77 agent-hours and 215.6 million tokens over 21.5 hours of wall-clock time, without human intervention. Wall-clock time is the time on the clock. Agent-hours add up each session’s time, so 77 can be larger than 21.5. The post’s 21 hours, roughly 950 agents, and 210 million tokens stay the post’s round numbers. The paper’s 21.5 hours, 949 sessions, and 215.6 million tokens stay the paper’s. The post’s “over 200,000,” “3,500,” and “20” stay the post’s. The paper’s 198,290, 3,564, and 19 reports stay the paper’s. Nineteen reports is the paper’s count of write-ups at the end. It is not the post’s 20.

The DNA on this card, as the preprint prints the moment. A worker agent read the DNA just upstream of one of these phage RTs. Upstream means the stretch of DNA before the gene. The paper quotes the trace: the L0050 flank, a 228,907-base-pair Logan metagenomic contig, is spectacular, and the agent can see a tandem repeat by eye. The two repeat words it prints are CATGTGTATCGCATGTT and CATGTGTTTCGCATGT, repeating with spacers of about 100 to 180 base pairs, and it calls that a CRISPR-like or msDNA-like repeat array. A base pair is one rung of the DNA ladder. A contig is a stretch of assembled sequence. Logan is the database name printed in that trace. The figure says that flank holds 14 copies of a 16-letter repeat, allowing at most one mismatch. The card image is that sequence block. Its header reads L0050_upstream_region. It is the DNA the paper says the agent was reading. It is not a photograph of the Bay Area lab.

What the preprint says the experiments do not yet show. The abstract calls ART a novel family of jumbo-phage RTs coupled to an array of about 200-nucleotide units and a dedicated partner gene. A nucleotide is one letter of DNA or RNA. The abstract says the arrays are highly expressed and show up as discrete units during infection by a Staphylococcus phage, which suggests an RT system steered by a set of distinct RNAs. Staphylococcus is a group of bacteria. The discussion says the units are transcribed in large excess over the enzyme, which suggests one enzyme supplied with a repertoire of templates or baits. Transcribed means copied into RNA. It then says they have not shown that the RT is active, and they have not shown that those unit RNAs are the enzyme’s substrates. A substrate is the molecule an enzyme acts on. Whether the RT and its partner interact, and what the system does for the phage, are currently unknown. Those limits are the preprint’s. The post’s “we don’t yet know its function” is the same limit in fewer words. This desk did not rerun the infection study or the database search.

Plain English for the rest of the card. ART is the name Anthropic gave this enzyme system: an RT, a partner gene, and a long repeat array. An RT copies RNA into DNA. CRISPR is the older repeat system that became a gene editor. A phage is a virus that infects bacteria. A tandem repeat is the same short DNA word copied down the line. Non-coding DNA is not a protein recipe. BSL-1 and BSL-2 are the lower lab grades. On this post, the lab does not handle human pathogens, and people do the bench work. A token is a chunk of text. About 950 agents, 21 hours, and 210 million tokens are the news post. 949 sessions, 21.5 hours, 77 agent-hours, and 215.6 million tokens are the preprint. San Francisco is the preprint’s affiliation line. Bay Area is the post’s line for the lab.

PRIMARY here: Anthropic’s 23 Sep 2026 news post, “Claude discovers a novel enzyme system with CRISPR-like repeats,” page date Sep 23, 2026, with no hour on the page this desk read — Tier A PRIMARY, the company’s own announcement — plus the same-day preprint PDF linked from that post. The group, the spring 2026 start, the Bay Area lab, the BSL-1 and BSL-2 limit, the no-human-pathogens line, the human-only wet lab, the high-level-direction line, the 21-hour campaign, the roughly 950 agents, the 210 million tokens, the one agent spotting the repeat, the three-part description, the earlier identification of the jumbo-phage RT, the short-RNA experiment, the unknown function, the handful-of-systems sentence, the over-200,000 / 3,500 / 20 census, the weeks-to-months line, the agent’s quoted exclamation, and the Feng Zhang quote are the post’s. The author list, the San Francisco affiliation, the 1.9 billion and 1.94 billion cluster counts, the 198,290 filtered RT clusters, the 3,564 families, the 949 sessions, the 77 agent-hours, the 215.6 million tokens, the 21.5 hours, the 119 tasks, the 19 reports, the L0050 contig, the two repeat words, the 14 copies of a 16-letter repeat, the Staphylococcus phage expression, and the explicit limits that the RT has not been shown to be active and that the system’s job for the phage is unknown are the preprint’s. NOT claimed: that ART edits genes, that the RT is active, that the short RNAs are the enzyme’s substrates, that the partner protein’s job is known, a named disease, a drug, an FDA filing, that this desk reran the search or the infection experiment, a stock tip, or investment advice. Distinct from the already-filed oracle-health-revenue-cycle-ai, enveda-311m, and ultrasight-24m.