TEHNOLOGIE
ART, the enzyme system found by Claude's agents — what was shown, what was not, and why RNA-based antivirals stay far off
Anthropic announced on 23 September 2026 that agents run by Claude surveyed 1.9 billion protein clusters on their own and surfaced ART, an undescribed family of enzymes whose DNA carries repeats reminiscent of CRISPR.
The stake is not the enzyme, whose function is still unknown, but what was shown about the models: a system of agents noticed an anomaly that no rule in its brief called for, questioned it on its own, and carried it through to a report. That touches directly the step in research long assumed to stay human — the curator's work, the close reading of raw data.
This article's testable thesis: the result is solid as method and fragile as biology. What was shown is that models can generate good hypotheses by reading primary data; what was not shown is that the enzyme does anything. The dominant scenario: ART stays an object of basic research, and the road to any medical application, including RNA-based antivirals, is measured in years and runs through stages the paper does not even touch.
What was announced, in figures
The paper, signed by Peter H. Yoon, Januka S. Athukoralage, Emmanuel Ameisen, Eric Kauderer-Abrams, Nicholas T. Perry and Matthew G. Durrant, is published as a preprint and has not been peer reviewed. The life sciences research group was formed in the spring of 2026, in the San Francisco area, and the laboratory works only at biosafety levels 1 and 2, with no human pathogens, according to the company's announcement.
The campaign started from a single research brief asking the agents to find new reverse transcriptase systems on the basis of partner-gene associations. The agents built their own search profiles, gathered about 198,000 clusters, sorted them into nine classes, scored 3,564 neighbouring protein families and promoted 17 candidates. In total: 119 tasks, 949 sessions, 77 agent-hours and 215.6 million tokens, over 21.5 hours of wall-clock time.
How the agent system worked
The structure will be recognisable to anyone who has built a multi-agent system. Each task is carried by a pair: a worker agent proposes a plan and executes it, a supervisor agent reviews its plan and results. The supervisor can open new tasks from what the worker observed, so the campaign extends itself as it finds things. A curator writes everything into a shared memory, and an editor reviews the final reports.
The number that matters here is not the count of agents but the proportion: 98 of the 119 tasks were opened along the way, from observations, not from the initial plan. The campaign ended with 19 reports, ranked against each other by a tournament-style evaluation. Of the 17 promoted families, only three were confirmed as new associations; 14 were rejected as annotation artefacts, parts of already described systems, or mere neighbourhood residents.
The moment of discovery: the agent read the DNA and saw the pattern
What makes the paper unusual is that the authors opened the session transcripts to reconstruct how ART was reached. The chain has four steps, each on the record.
First, the census promoted an unknown neighbouring gene. Then a worker rejected it: it was a subunit of a phage polymerase, so the "accessory" reading collapsed. But the same agent noted a secondary finding — a jumbo phage carried a free-standing retron-like enzyme — and queued a follow-up task. The supervisor set the next worker to check the upstream region, where retrons keep their RNA. The enzymes beside the polymerase had only 27 base pairs there, too few; their close relatives had a median of 940.
The fourth step is the decisive one: the worker loaded the raw sequence into its context and read it as text. The transcript records the agent's reaction, describing the repeated pattern seen "by eye" and calling it on the spot "a CRISPR-like repeat array". Immediately after, the same transcript shows the agent questioning its own find and calling for a novelty test against the literature — the step that usually separates a researcher from a hypothesis generator.
What ART is, in plain words
The system has three pieces. The first is a reverse transcriptase, an enzyme that copies RNA into DNA; the family is old and well known, and variants of it already sit inside laboratory tools. The second is an array of non-coding DNA, that is a short sequence repeated several times, with different stretches between the repeats. The third is a partner gene, immediately after the enzyme.
This combination is described nowhere in the literature, and that is the paper's claim: the enzyme is not new, the arrangement around it is. The searches found 95 distinct clusters in jumbo phages, 28 of them with a detectable repeat array. The arrays hold between 3 and 21 repeats, and the enzyme has an unusually long extension at one end, about 180 amino acid residues, against 50 or fewer in its relatives.
The resemblance to CRISPR is one of form, not function, and the paper says so explicitly. The technical text compares ART rather with retrons, bacterial anti-phage defence systems described in the specialist literature, which also have three pieces but a single RNA. In CRISPR, the repeats are near-identical and alternate with short stretches of about 30 bases that are gained and lost between related strains. In ART, the stretches between repeats are four to seven times longer and are conserved between related phages. On top of that, no ART site has the genes that make CRISPR a tool.
What was actually verified in the laboratory
This needs careful reading, because it is the only part with living cells. The people in the laboratory did two things. They took public RNA data from an already published study, on a staphylococcus phage that carries an ART system, and saw that the array is heavily transcribed at every stage of infection — at 15 minutes it reaches up to 8% of all the phage's RNA, among the most abundant transcripts. Then they put the system on plasmids in a laboratory bacterium and found the same short, distinct RNAs.
That is all. It was not shown that the enzyme is active. It was not shown that those RNAs are its substrate. It was not shown that the enzyme and the partner actually touch. What it does for the phage is unknown. The paper itself lists these four unknowns, in a paragraph the press releases moved past faster.
The most honest part: ten reruns, zero rediscoveries
The authors ran the same campaign ten more times. Nearly all of them reached ART loci in the census, and in two cases the agents even investigated the family as a follow-up. None read the upstream DNA, so the array was missed in every rerun.
Acknowledging this changes how the whole announcement should be read. The discovery is not a reproducible output of a method but structured luck: a system that runs enough threads, with enough freedom, that now and then one of them takes the right step. The difference between those two things is enormous for anyone budgeting research on the strength of it.
To isolate the capability anyway, the authors built a fixed-input benchmark with five levels of information, in which the model is asked to describe the system and a judge model scores the report against ten features. This is where the most practically useful result appears.
A result about models, not only about biology
On the benchmark, the four more capable models pulled clearly away from the other three. But the interesting finding is a different one: the tools got in the way. When the system was given directly in context, the good models described the array correctly in over 90% of attempts; when it was given as files, with analysis tools available, the rate fell to as low as 32%.
The explanation, checked in the transcripts, is that the models with tools stopped reading the sequence. In 39% of the attempts with files, the model never brought a continuous stretch of even 200 bases into context, meaning it never saw more than one repeat. When it did read that much, recognition rose by 16 to 32 percentage points.
For anyone building agent systems, that is the transferable lesson and it has nothing to do with biology: an agent that has tools tends to delegate the observation to the tool and stop looking at the data itself. Tools raise what an agent can do and lower what it sees. The same asymmetry appears in any pipeline where a model calls a search function instead of reading the rows.
The actors and what each said
| Who | Role | Position, per the sources |
|---|---|---|
| Dario Amodei | chief executive, Anthropic | "may constitute a new gene editing mechanism that could have applications in gene therapy"; cited the discovery at the UN Security Council |
| Feng Zhang | MIT and Broad Institute, CRISPR pioneer | "an exciting example of how AI agents can contribute to biological discovery"; the association "merits further investigation" |
| Stanley Qi | bioengineering, Stanford | AI could greatly expand the capacity to explore, "in this case, in just 21 hours" |
| Kevin Blake | microbiologist, Washington University | "There's nothing to indicate this is a rival to CRISPR-the-technology, or could be developed into any kind of therapeutic" |
| The paper's authors | Anthropic, San Francisco | state explicitly that the enzyme's activity, the role of the RNAs and the biological function remain undemonstrated |
The context of Amodei's statement matters. He spoke on 23 September, by video link, at the UN Security Council, where he and Sam Altman called for international standards on artificial intelligence, according to CNN. The discovery was used there to illustrate the upside, inside a speech about risk — a double posture that some commentators also noted.
RNA-based antivirals: what already exists and how far ART is
The natural question after a headline with "CRISPR" in it is whether a treatment comes out of this. Three things that the press mixes together are worth separating.
First: tools that cut viral RNA already exist and have been published for years. The approach called PAC-MAN, described in Cell in 2020, uses an enzyme from the Cas13 family to degrade coronavirus and influenza RNA in cultured human lung cells. It is a six-year-old proof of principle which, for all the pandemic interest, did not become a medicine. The obstacles are not about recognising the target but about delivery into the body, immune reaction and safety.
Second: reverse transcriptases are already parts inside editing tools. The precision editing described in Nature in 2019 ties such an enzyme to a Cas9-type protein and writes the change from an RNA template. So the fact that ART contains a reverse transcriptase is not, in itself, news for the tools field.
Third, and the one that is missing: for ART to mean anything therapeutic, it would first have to be shown that the enzyme is active, then what it does, then that it does that in human cells, then that it can be delivered and tolerated, and then would come the years of clinical trials. The paper has not yet cleared the first stage. Between "we found an undescribed arrangement in a virus of bacteria" and "antiviral" there are five gates, and the first one is not open.
What stays valuable regardless of applications: if ART really is a new class of non-coding repeat element, it is a piece of basic biology that needs no application to be worth studying. The paper also notes that another group recently found arrays of RNAs beside an unrelated family of enzymes, using a language model built specifically for non-coding elements — a sign that the architecture arose more than once in nature and is not an isolated case.
The counter-hypothesis: how much is discovery and how much is communication
The objection from microbiologist Kevin Blake, made to Al Jazeera, is the most serious and deserves facing head-on: there are countless CRISPR-like sequences still uncatalogued, because millions of bacterial species have not been studied. In other words, in a database of 1.9 billion protein clusters, finding something undescribed is expected, not exceptional. What is not expected is the way it was found.
There is a sharper objection too, which the paper defuses by itself. The reading "AI discovers a new CRISPR" rests on a resemblance of form that the technical text rejects, listing three differences from CRISPR and the absence of the associated genes. The distance between the preprint and the press release is, in this case, the distance between "reminiscent architecture" and "gene editing mechanism".
What would settle it: an independent paper showing the enzyme's activity. Until then, the honest assessment is that we have a good methodological result and an unfinished biological discovery, announced together at a politically chosen moment.
What it means for Romania
The useful part for a reader in Romania is not the enzyme but the fact that the raw material is public. The database the agents ran on is built from public assemblies of the world sequence archive; the Logan project, which turned 44.1 petabases of raw data into 0.9 petabases of assembled sequence, is open to anyone. The database is not the barrier.
The barrier is twofold and of another nature. The first is compute cost: 215.6 million tokens for a single campaign is not a sum that slips into a small grant, and the reruns show that one campaign is not enough. The second is the laboratory: the part that gives the result its value was done by people, with cells, and there a country without access to such facilities stays a consumer of preprints.
The practical conclusion for a research group here is that sequence mining has become an activity where a small team can compete, provided it has a compute budget and a partner able to test at the bench. It is, for the first time, an area where the lack of large sequencing infrastructure is no longer decisive at the first step.
Scenarios
Dominant scenario: ART stays basic biology. Trigger: characterisation takes time, and the function turns out to be phage defence, not something programmable. Constraint: nothing yet shows the enzyme is active. Consequence: a few follow-up papers, no product, and the announcement stays cited mainly as a demonstration of method. Indicator: no paper showing enzymatic activity appears by the end of 2027.
Plausible scenario: the function is clarified and turns out interesting. Trigger: an independent laboratory shows what the short RNAs and the partner do. Constraint: even so, moving into human cells is a different piece of research. Consequence: ART enters the textbooks of phage biology, and competition on agentic mining accelerates. Indicator: a peer-reviewed paper signed by a group outside Anthropic.
Tail scenario: the method produces something with a direct application. Trigger: similar campaigns, run by several teams, surface a system with demonstrable and useful activity. Constraint: the reruns show the success rate per campaign is low. Consequence: agentic mining becomes a separate funding line in biotechnology. Indicator: at least three different groups publish discoveries obtained with autonomous agents in the next eighteen months.
Testable predictions
| Prediction | Horizon | Probability | How it is checked | Verification source |
|---|---|---|---|---|
| The ART preprint appears in a peer-reviewed journal | by 31.12.2027 | 45–60% | Published article, with the same authors | PubMed, the journal's site |
| Enzymatic activity of an ART reverse transcriptase is published | by 31.12.2027 | 25–40% | Paper with biochemical activity data | bioRxiv, PubMed |
| A group outside Anthropic publishes a paper on ART | by 31.12.2027 | 35–50% | Paper with authors holding no Anthropic affiliation | PubMed, Google Scholar |
| ART is described in a publication as a working gene-editing tool | by 31.12.2028 | 5–12% | Paper showing directed editing with ART components | PubMed |
| At least three other groups announce molecular systems found with autonomous agents | by 31.03.2028 | dominant scenario | Three distinct announcements, with preprint or article | bioRxiv, institutional releases |
| An antiviral based on cutting RNA enters phase 1 trials in the EU or the US | by 31.12.2028 | 15–30% | Registration in a clinical trials registry | ClinicalTrials.gov, the EU registry |
Method, definitions and figures
The terms. A "reverse transcriptase" is an enzyme that copies RNA into DNA. A "repeat array" means a short sequence that recurs, with different stretches between the repeats. A "phage" is a virus that infects bacteria, not people; a "jumbo phage" is a category defined by genome size. A "preprint" is a text published before peer review.
The campaign figures. 1.9 billion clusters, about 198,000 clusters retained, nine classes, 3,564 neighbouring families, 17 candidates, 119 tasks, 949 sessions, 77 agent-hours, 215.6 million tokens and 21.5 hours of wall-clock time all come from the preprint. The press rounded these to "950 agents", "210 million tokens" and "21 hours"; this article uses the paper's figures.
The ART figures. 95 clusters, 28 with a detectable array, between 3 and 21 repeats, the extension of about 180 amino acid residues and the share of 8% of the phage's RNA at 15 minutes all come from the preprint. The 8% share comes from reanalysing public data from an earlier study, not from a new experiment.
What was not independently verified. None of the paper's biology can be checked from outside a laboratory; this article checks the public claims against one another and against the technical text, not at the bench. The comparisons with Cas13 and with precision editing rest on papers published in 2020 and 2019 respectively, cited above.
The probabilities. The bands in the table are editorial estimates, in ranges of at least seven percentage points, not calibrated on a track record. They show the order of magnitude of confidence, not a measured frequency.
Limits of the analysis
The analysis rests on the Anthropic preprint, on the company's announcement and on press reporting from 23–24 September 2026. The preprint has not been peer reviewed, and the raw data and full session transcripts are not public, so the reconstruction of the discovery relies on the excerpts the authors chose.
What the article cannot establish: whether the enzyme is active, what it does for the phage, what the campaign cost in money, and whether the benchmark results transfer to other protein families. The claims about the model's internal signals cannot be checked either, as that would require access to the model.
The diagnosis would be invalidated by a paper showing enzymatic activity and a clear function, which would move ART from "promising hypothesis" to characterised system, or, conversely, by an independent reanalysis showing the repeat array to be an assembly artefact. The natural point for reassessment is the first paper signed from outside the company.
.jpg?width=1200)
