Claude Found a Phage Pattern. Ten Repeat Searches Missed It

Anthropic researchers report that a network of Claude agents surfaced a previously uncharacterized pattern around reverse‑transcriptase genes in jumbo phages: a long array of repeated DNA beside the enzyme and a partner gene. Human scientists then found that the array is expressed as several short RNAs. That is a biological result worth examining, but it is narrower than the comparison with CRISPR in [Anthropic's September 23 announcement](https://www.anthropic.com/news/claude‑discovers‑novel‑enzyme‑system). The accompanying [40‑page preprint](https://www‑cdn.anthropic.com/22573675ada52a8ca8a97a1a4b4326b2f208a071.pdf) does not show that the reverse transcriptase is active, that the short RNAs are its substrates, that the enzyme and partner protein interact, or what the system does for the phage. It presents an early biological finding and an unusually detailed account of how an AI‑assisted search succeeded once, then missed its defining observation in ten repeat campaigns. The underlying reverse‑transcriptase lineage had appeared in earlier phage‑genome work. The new observation was the arrangement around it. The authors named the family array‑associated reverse transcriptases, or ART. Their analysis reports 95 ART reverse‑transcriptase clusters at 90% identity. Twenty‑eight had a detectable upstream array containing three to 21 repeat copies. The researchers found no nearby `cas` genes, and the spacer organization differs from a typical CRISPR array. In this evidence, “CRISPR‑like” describes a visual genomic resemblance rather than a demonstrated CRISPR mechanism. As of October 9, no independent laboratory validation is identified in the sources reviewed for this analysis. The public discovery registry [What AI Found classifies ART as a claimed, collaborative result](https://whataifound.org/finding/2026‑09‑23‑array‑associated‑reverse‑transcriptases) and notes that nobody outside the lab has checked it. The preprint's enzyme activity, substrate, partner interaction and biological function therefore remain open research questions.

The discovery began with a human‑written brief asking agents to find new reverse‑transcriptase systems through previously unknown partner‑gene associations. Anthropic's harness split that job into tasks, pairing a Claude Code worker that proposed and ran an analysis with a supervising agent that reviewed the plan and results. Agents could open follow‑up work and wrote their output to a shared record. Their endpoint was a set of reports for human reviewers. Running Claude Mythos 5, the campaign searched a database containing about 1.9 billion protein clusters. It recovered roughly 200,000 reverse‑transcriptase clusters, sampled about 11,000 loci and scored 3,564 recurring neighboring protein families. Sixteen candidate families passed the campaign's filters, and an agent‑generated follow‑up added a seventeenth. The complete run comprised 119 tasks, 949 agent sessions, 215.6 million tokens and 21.5 hours of wall‑clock time. The paper calls this period free of human intervention, which applies to the campaign after researchers supplied its brief, harness and data rather than to the project as a whole. Most promoted candidates did not survive review. Three of the 17 candidate partner families were retained as previously unreported associations; the other 14 were rejected or set aside. ART emerged from a side path rather than the partner‑family ranking itself. A worker rejected one apparent neighboring‑gene association, opened a closer look at the reverse transcriptase, read the raw upstream DNA and recognized the tandem repeats. It counted the repeats, compared the layout with known systems and searched the literature before filing a report for people to assess. That division of labor matters when describing autonomous discovery. The agent made the anomalous observation and pursued it computationally. People defined the broad question, built the environment, selected work for follow‑up and performed every laboratory experiment. The evidence supports AI‑directed anomaly detection inside a bounded research system; it does not describe an AI independently choosing a field, acquiring samples or operating a lab.

The paper's strongest biological evidence comes from RNA measurements. A Claude Science session found an existing time‑course dataset for Staphylococcus phage SA1, which carries ART. Reanalysis showed array‑derived RNAs at several stages of infection; 15 minutes after infection they accounted for as much as 8% of phage RNA and appeared as shorter species with repeatable boundaries. Anthropic's human scientists then expressed the SA1 ART region on plasmids in *E. coli* and performed small‑RNA sequencing. They again observed discrete short RNAs from the array under both its native locus and a separate promoter. This supports the claim that an ART array can produce several short non‑coding RNAs. It does not establish what those RNAs bind, whether the reverse transcriptase copies them, or whether the system helps a phage compete with another phage. Computational predictions sit beside those experimental observations in the paper. The association of ART with three different partner‑protein families comes from sequence searches and predicted structures. A proposed interface between the enzyme and partner comes from structure prediction with moderate confidence. RNA expression comes from sequencing data and a human‑run laboratory experiment. Enzyme activity, RNA substrate identity, protein interaction and biological function remain unshown.

Anthropic reran the same broad genome‑mining campaign ten times to test whether it would rediscover ART. Nearly every campaign that completed the census sampled ART loci, and two workers investigated the lineage as follow‑up. None read the DNA upstream of the reverse transcriptase. All ten therefore missed the repeat array that defined the original discovery. The authors attribute the result to the broad protein search space and the harness's non‑deterministic behavior. The failure is more informative than a simple one‑in‑eleven success rate because the paper traces a specific bottleneck. In a separate fixed‑input benchmark, the researchers gave models ART sequences directly in context or as files with tools. A Mythos 5 judge model then scored the reports against ten author‑selected features. The four strongest Claude models described the array in at least 90% of attempts when the relevant loci were placed directly in their context. Performance fell as low as 32% in a tool‑using setup. Inspection of those four models' runs showed that 39% of file‑based attempts never read a continuous stretch of at least 200 nucleotides, so the model did not see more than about one repeat unit. Reading at least 200 nucleotides raised recognition by 16 to 32 percentage points for each of those four models. Across the four models, recognition rose from 29% to as high as 76% as more DNA entered context, reaching as high as 96% for Mythos 5. The practical lesson is about information exposure: an agent can possess search tools and the correct files yet fail because it never brings the decisive evidence into view. Those percentages have their own limit. They come from fixed inputs, an internal judge model and criteria selected by the paper's authors. They help explain the original miss, but they are not an independent replication of ART or a general measure of autonomous scientific ability. The ten end‑to‑end failures show that recognizing a pattern after it enters context is different from reliably finding that pattern inside an open‑ended search.

The ART work establishes a coherent genomic pattern and experimentally confirms its unusual RNA output. It also records a model noticing a feature that earlier annotations had missed. That combination makes the research more substantive than an untested model‑generated hypothesis while leaving the central biological question unanswered. Evidence that would change the interpretation is concrete: demonstrate reverse‑transcriptase activity, identify its substrate and product, verify an interaction with the partner protein, determine what the system does during phage infection, and reproduce the findings outside Anthropic. Peer review and independent laboratory work can then test whether ART is a useful biological system or simply an interesting arrangement. Until then, the defensible result is that Claude helped researchers find where to look, not that it discovered a new gene‑editing or medical technology. *Figure: The original ART campaign found the repeat array after reading raw DNA, while ten repeat campaigns missed it. In the paper's fixed‑input breakdown of the four strongest models, 39% of file‑based attempts never read 200 contiguous nucleotides; reading at least that much raised recognition by 16–32 percentage points for each model. Source: [Yoon et al., Anthropic preprint, Figure 4](https://www‑cdn.anthropic.com/22573675ada52a8ca8a97a1a4b4326b2f208a071.pdf). Figure by OpenTools Team; no third‑party expressive material used.*