The short version: On September 23, 2026, Anthropic said a team of Claude agents had found a new family of enzymes, called array-associated reverse transcriptases (ART), sitting beside CRISPR-like repeats in the DNA of viruses that infect bacteria. The preprint behind the announcement shows that the discovery was not the assignment. The agents had been told to look for new partner genes. One of them, reading raw DNA next to an enzyme, noticed a repeating pattern nobody had asked about and flagged it. The campaign ran 21.5 hours without human intervention, 14 of its 17 leads did not hold up, and human scientists did all the lab work. For a business owner, the useful part is not the biology. It is that the most valuable output was the flag outside the question.
What did Claude discover?
According to Anthropic’s announcement and the preprint it links, ART is a family of reverse transcriptases, enzymes that copy RNA into DNA, found in very large bacteriophages. Each one comes packaged with a dedicated partner gene and an array of repeating DNA units of roughly 200 letters each. The arrays resemble CRISPR arrays, but no CRISPR-associated genes appear anywhere near them.
The arrays are also busy. Using a published infection study of a Staphylococcus phage, the agents found them highly expressed at every stage of infection, appearing as distinct units, which suggests one enzyme directed by a whole set of different RNAs. What the system actually does is still unknown, and Anthropic says experiments are ongoing. The underlying enzyme had turned up in earlier studies; by Anthropic’s account, Claude appears to be the first to recognize the repeats and the partner gene that define the family.
Feng Zhang of MIT and the Broad Institute, a pioneer of CRISPR gene editing, is quoted in Anthropic’s announcement calling it “an exciting example of how AI agents can contribute to biological discovery.” He is not an author of the preprint; the six authors are Anthropic scientists, and Anthropic presents the work as the first result from its new life sciences research group and laboratory.
How did the Claude agents actually work?
The preprint describes a small organization made of Claude instances, running on Claude Mythos 5. A worker agent plans a task and runs it. A supervisor agent reviews the plan and the result, accepts it or sends it back, and can open new tasks when something interesting turns up. A curator writes findings into a shared knowledge base that later agents read. An editor reviews each report before it is filed.
The scale is striking. The agents built their own search models, pulled about 200,000 reverse transcriptase clusters out of 1.9 billion protein clusters, sorted them into nine classes, sampled about 11,000 of their DNA neighborhoods and scored 3,564 candidate partner families. Sixteen passed the agents’ own criteria and a seventeenth was promoted later. In all, the campaign ran 119 tasks and 949 agent sessions, 76.9 agent-hours and 215.6 million tokens, over 21.5 hours of wall-clock time.
The supervision was real work, not decoration. Supervisors opened 98 of the 119 tasks as follow-ups, and 49 tasks were revised at least once after review.
Where did the discovery come from?
Not from the question. The research brief asked for new partner genes next to reverse transcriptases. The find came when a worker loaded the DNA beside one phage enzyme, read it as text and wrote, in a transcript the preprint reproduces, “I can see by eye a tandem repeat array.” It named what it saw, and the supervisor structure turned that aside into follow-up tasks.
The preprint makes the point directly: in earlier AI-assisted discovery, a human set the expectation of what normal looks like and judged what was new. Here, the agents supplied both, drawing on what they already knew rather than on what their task told them to find.
How many of the leads held up?
Fewer than one in five. Of the 17 candidate families, the campaign confirmed only three as previously unreported. The other 14 were rejected or set aside as annotation errors, parts of systems already described, or genes that simply lived nearby without working with the enzyme. Separately, flags on individual enzymes produced three new lineages.
Then the humans took over. Anthropic says its scientists did all the hands-on work, expressing proteins and characterizing them in a lab limited to biosafety levels 1 and 2, with no human pathogens, and decided which candidates deserved an experiment. The AI narrowed a haystack; people tested the needles.
Which Claude models can do this?
The preprint also tested seven Claude models on characterizing the ART system from the same inputs, with each report scored by Mythos 5 against ten curated features. It found a clear gap. Opus 5.5, Mythos 5.1, Mythos 5 and Opus 5 formed the top group; Opus 4.6, Opus 4.8 and Sonnet 5 fell behind. One caution: the judge was itself a Claude model. Still, it is worth knowing that Opus 5.5, released September 22 at a lower price, landed in the top group.
What can a small business take from a biology paper?
We have written before about when it is safe to stop watching an AI, and the answer there was a checker that rejects every wrong step. This paper adds a different lesson: the best result came from something the agent noticed, not something it was asked.
Most owners use AI the way the research brief was written: summarize these invoices, sort these reviews, draft this quote. Add one line to the request: and tell me anything that looks unusual, even if I didn’t ask about it. The vendor who bills a little more every month, the customer who went quiet after years of steady orders, the same complaint coming back in three reviews. Those are the tandem repeats in a small business’s data.
Then keep the ratio in mind. Here, 14 of 17 leads did not hold up. Most of what an AI flags as odd will turn out to be nothing. Your job is the bench work: checking the few flags that survive, which is also the part of the job that was always yours.
Frequently Asked Questions
What did Claude discover?
Claude agents identified array-associated reverse transcriptases, or ART, a new family of enzymes in large bacteriophages that pair with a dedicated partner gene and arrays of repeating DNA units resembling CRISPR arrays. Anthropic announced the finding on September 23, 2026, and says the system’s function is still being studied.
Did Claude do the lab experiments?
No. The agents ran the database searches, the analysis and the literature review for 21.5 hours without human intervention. Anthropic says its human scientists did all the hands-on lab work in a biosafety level 1 and 2 lab and decided which candidates were worth testing.
How accurate were the Claude agents?
Of 17 candidate enzyme families the agents investigated, only three were confirmed as previously unreported; the other 14 were rejected or set aside. Supervisor agents also reviewed every task, and 49 of 119 were revised at least once.
Which Claude models performed best in the study?
In the preprint’s follow-up test of seven Claude models, Opus 5.5, Mythos 5.1, Mythos 5 and Opus 5 formed the top group, while Opus 4.6, Opus 4.8 and Sonnet 5 scored lower. The reports were scored by Mythos 5, itself a Claude model.
If you asked an AI to tell you what looks unusual in your own books this month, what do you think it would find?
