Anthropic reports that autonomous Claude agents surfaced ART, a repeat-associated reverse-transcriptase system in phages, but the preprint also exposes important limits in reproducibility and biological validation.
Anthropic has reported an unusual result from its new biology lab: a large collection of Claude agents searching genomic data identified a previously uncharacterized arrangement of reverse transcriptases, neighboring genes, and tandem DNA repeats in bacteriophages. The researchers call the family array-associated reverse transcriptases, or ART.
The result is scientifically interesting for two separate reasons. The biological system itself appears to be new, although its function remains unknown. The route to finding it is also unusual: according to Anthropic's September 23 report, the candidate emerged from an autonomous computational search rather than from a narrowly specified pattern detector.
This is not evidence that an AI independently solved the biology of a new CRISPR-like system. The underlying work is a preprint, not a peer-reviewed study. Human scientists performed the laboratory experiments, ART's reverse-transcriptase activity has not yet been demonstrated, and the researchers do not know what the system does. The search also has a striking reproducibility limitation: repeated agent runs did not rediscover the tandem-repeat feature.
Those caveats make the result more useful, not less. ART is a concrete test case for what scientific agents can currently contribute to discovery and where the difficult parts still remain.
The scientific problem is familiar in genome mining. Public sequence databases contain enormous amounts of genetic material, including proteins and genomic neighborhoods that have never been experimentally characterized. Finding unusual combinations can reveal new molecular systems, but deciding which anomalies deserve attention is labor-intensive.
In the preprint, Autonomous AI agents discover reverse transcriptases with tandem repeat arrays, Peter H. Yoon and colleagues used Claude Code agents to survey reverse-transcriptase loci across roughly 1.9 billion protein clusters. Reverse transcriptases, or RTs, copy RNA into DNA and occur in a wide range of mobile elements and microbial defense systems.
Anthropic reports that the campaign involved about 950 agent sessions over roughly 21 hours and consumed about 210 million tokens. The agents collected more than 200,000 RTs, identified thousands of candidate systems, and produced detailed reports on a much smaller set selected for closer inspection.
One agent examining the DNA around an unusual RT noticed a tandem array of regularly spaced sequences. It then characterized the repeats, compared the locus with known RT systems, searched the literature, and escalated the candidate for human review.
That last step is important. The system was not found because researchers explicitly asked a model to detect this particular repeat architecture. The agent was investigating a broader RT-search problem and shifted its attention from proteins to nearby non-coding DNA.
The reported ART loci occur mainly in bacteriophages and combine three conspicuous features: an RT, a neighboring partner gene, and an array of repeated DNA units. The RT family itself was not wholly unknown; related RTs had appeared in previous work. The new claim is that these RTs belong to a broader molecular system defined by the associated repeat array and partner protein.
The repeat architecture invites comparison with CRISPR because CRISPR systems also contain arrays that are transcribed into multiple RNAs. But the resemblance is structural, not functional evidence.
The researchers found that ART arrays are expressed as discrete RNAs. Analysis of existing phage-infection transcriptomic data showed strong expression from an ART array, and follow-up experiments by the Anthropic team produced multiple short RNA species from the array. This moves ART beyond a sequence pattern found only in silico.
It does not establish what those RNAs do. The study has not shown that the ART RT acts on them, that the neighboring protein interacts with the RT, or that ART performs sequence-programmable cutting, copying, or integration. The ART loci also lack the Cas genes that define CRISPR-Cas systems.
Calling ART a new gene-editing system would therefore outrun the evidence. At present, it is a newly described genomic architecture with experimentally observed RNA expression and an unresolved molecular function.
The preprint contains an unusually useful stress test of the discovery story. After ART had been found, the researchers repeated the broader agent search ten times using the same task and harness. None of those runs independently rediscovered the tandem-repeat feature.
This does not invalidate the biological observation. The genomic loci still exist, and the RNA-expression evidence is separate from whether an agent can find the pattern again from a broad search space. It does, however, constrain the stronger claim that this workflow is already a reliable discovery engine.
The repeated runs often reached relevant RT candidates, but the agents did not inspect the upstream DNA in the way that produced the original discovery. The bottleneck was therefore not simply recognizing a tandem repeat once it was presented. It was choosing, among many possible follow-up analyses, to look at the right sequence at the right time.
That distinction matters for scientific-agent evaluation. A model can possess the local capability needed to recognize an interesting pattern while an autonomous search process still has low probability of navigating to it. Scientific discovery involves both inference and allocation of attention.
For reproducibility, future systems will need evaluation methods that measure this search behavior across repeated runs, not only showcase successful trajectories. Reporting unsuccessful campaigns, candidate funnels, computational cost, and intervention points would make comparisons much more informative.
The phrase "autonomous discovery" also needs a precise boundary. Anthropic says its scientists supplied a high-level research prompt while the agents performed much of the computational exploration and candidate analysis. Human researchers then reviewed candidates and performed the laboratory work.
That division is scientifically sensible. Genome mining can generate large numbers of plausible stories from sequence correlations, and experimental validation is what separates an interesting computational pattern from a characterized biological mechanism.
Here the experimental evidence remains early. RNA expression supports the idea that the repeat array is biologically active, but the central mechanistic questions are unanswered. Demonstrating RT activity, identifying its substrate, establishing the role of the partner protein, and determining whether different array RNAs direct distinct reactions would substantially strengthen the biological case.
Independent replication would strengthen it further. The current evidence comes from the same organization that developed the agent system and operates the laboratory reporting the discovery. That does not diminish the data, but it makes external confirmation particularly valuable before broader claims are made about either ART biology or the general effectiveness of the discovery workflow.
ART is a stronger demonstration of AI-assisted science than a benchmark showing that a model can answer biology questions. The system participated in an open-ended search over real genomic data, surfaced an anomaly that researchers considered worth investigating, and handed off a hypothesis that could be tested experimentally.
At the same time, the evidence draws a useful line around the achievement. The agents did not determine ART's function. The discovery trajectory was not reproduced in ten additional broad searches. The biological experiments were performed by people, and the work has not yet passed peer review.
The next important results will therefore come from two directions. Biologically, researchers need to determine what ART actually does and whether its repertoire of RNAs has a functional relationship with the RT. Computationally, scientific-agent systems need to show that they can repeatedly navigate large hypothesis spaces and surface valuable anomalies without relying on a lucky exploratory trajectory.
If both problems can be solved, genome mining is a natural domain in which agentic AI could become genuinely useful: sequence databases are vast, candidate generation is cheap relative to wet-lab validation, and every computational hypothesis eventually faces an experimental test. ART is an intriguing early example of that workflow, not yet proof that autonomous scientific discovery has become routine.
