"Autonomous AI agents discover reverse transcriptases with tandem repeat arrays" is a preprint released by Anthropic on September 23, 2026, credited to Peter H. Yoon, Januka S. Athukoralage, Emmanuel Ameisen, Eric Kauderer-Abrams, Nicholas T. Perry and Matthew G. Durrant, all listed at Anthropic in San Francisco, with Perry and Durrant as corresponding authors (Source: www-cdn.anthropic.com). It reports a genome-mining campaign in which instances of Claude Code, configured with Claude Mythos 5, surveyed reverse transcriptase (RT) loci across 1.9 billion metagenomic protein clusters and identified a family the authors name array-associated reverse transcriptases (ART). The document has not been peer reviewed; Anthropic's accompanying announcement calls it both a "pre-print" and a "technical report" (Source: anthropic.com).
The paper makes two kinds of claim that are separable: a biological one about a previously undescribed family of enzymes and its associated repeat arrays, and a claim about AI systems — that a general-purpose model reading raw DNA in its context noticed an anomaly outside the task it had been given, and that this recognition is traceable to internal signals in the model that respond to repeated sequence. The authors state explicitly what the work does not establish: "we have not shown that the RT is active or that the unit RNAs are its substrates," and "[w]hether the RT and its partner interact, and what the system does for the phage, are currently unknown."
The campaign and its accounting
The agents ran inside what the paper calls an agentic harness: a research brief was converted into five stages (input assembly, database sweep, RT classification, neighborhood census, deep dives), each stage closed by scripted completion checks and broken into tasks. Each task paired a worker agent, which wrote and executed a plan, with a supervisor agent, which reviewed the plan and results and could open follow-up tasks; a curator agent entered findings into a shared knowledge base read by later agents, and an editor agent reviewed reports before filing. Agents acted through the tools of Claude Code — a shell, file access and web search — with added connectors to a metagenomic protein database, literature retrieval, the knowledge base and a GPU job queue. The harness permitted up to 58 concurrent sessions.
The campaign accounting is given in the Results and Methods:
| Measure | Value |
|---|---|
| Protein clusters searched | 1,939,242,578 clusters at 50% identity (reported in the text as 1.9 billion), derived from approximately 15.8 billion proteins spanning about 11 million biosamples |
| RT clusters recovered | 198,290, from 203,381 retained sequences out of 3,391,714 candidate hits |
| RT classes | 9 (seven labelled classes, one "novel" class of 25,737 clusters, one "unplaced" class of 137,385) |
| Loci sampled in the neighborhood census | 10,983 loci from 7,238 of 7,308 anchor clusters |
| Recurring neighbor families scored | 3,564 |
| Families promoted to deep dives | 16, plus a seventeenth promoted later by a follow-up task |
| Reports filed | 19 (16 covering the 17 families, three on new RT lineages) |
| Tasks | 119 total: 107 completed, 10 rejected at triage, 2 stalled; 98 were follow-up tasks opened during the work; 49 were revised at least once |
| Agent sessions | 949 (one launch, 414 worker, 375 supervisor, 107 curator, 52 editor) |
| Agent-hours | 76.9, of which 63.7 in worker sessions |
| Tokens | 215.6 million (11.3M uncached input, 14.9M output, 189.5M written to the prompt cache; cache reads excluded) |
| Wall-clock time | 21.5 hours, without human intervention |
| Recorded tool use | 7,578 shell commands, 696 database queries, 131 literature searches, 61 web requests |
Of the 17 promoted partner families, three were confirmed as previously unreported RT associations and 14 were rejected or set aside as annotation artifacts, parts of already-described systems, or general neighborhood residents. Three further RTs flagged by workers were identified as new RT lineages. The 19 reports were ranked in a tournament of 342 pairwise comparisons judged by a Mythos 5 model against a fixed rubric weighting impact 0.35, novelty 0.30, soundness 0.25 and actionability 0.10, with ratings fitted by a Bradley–Terry model; the ART report ranked third, winning 32 of its 36 games.
How the array was noticed
The paper reconstructs the path to ART from the session transcripts, and states that the finding fell outside the brief: the campaign was directed at partner protein-coding genes rather than non-coding features, and "[n]either this task brief nor the research brief mentioned repeats or arrays." The candidate was first selected because the RT co-occurred with the β-like subunit of a jumbo-phage non-virion RNA polymerase. A worker rejected that association as spurious but queued a follow-up task on the RT itself; because retron RTs act with a non-coding RNA encoded upstream, a supervisor asked the next worker to check the 5′ flanks. The RTs beside the polymerase gene had short upstream regions (median 27 bp) while their closest relatives had long ones (median 940 bp), and the worker read the upstream sequence of those relatives directly into its context. The paper quotes the resulting reasoning trace, in which the worker writes that it can "see by eye a tandem repeat array," names the repeat pattern, and asks whether it is "a CRISPR mini-array or a retron-like repeat region." A subsequent trace shows the same worker questioning its own novelty claim and resolving to "characterize quantitatively, then run the novelty kill-test literature search with precise descriptors."
The Methods record checks on this account: the two retrievals before recognition returned sequence without annotation, no tool call intervened between them and the recognition, no earlier tool call contained a repeat analysis, and no established repeat finder was run in the session. Every repeat the worker named was recounted by the authors in the retrieved sequences, and the first pattern the worker wrote was recounted on both strands of the 2,900-nt sequence it had retrieved with at most one mismatch.
The ART family as described
Profile searches of the metagenomic database and public genomes identified 95 distinct RT clusters at 90% identity in cultured jumbo phages and predicted viral contigs, 28 of which carried a detectable repeat array upstream of the RT. Phylogenetic analysis placed the RTs in a clade beside retrons. The paper describes three defining features: an array of non-coding DNA repeats upstream of the RT, an unusually long N-terminal extension of about 180 residues where other RTs typically carry about 50 or fewer, and a dedicated partner gene directly downstream. The catalytic YxDD motif is retained in all 93 members whose sequences span it, the other two being truncated before it.
The 28 arrays adjacent to an RT span 0.3 to 4.1 kb and hold 3 to 21 copies of a short repeat at near-constant spacing; repeats are 15 to 49 nt long with a palindromic core of about 15 nt, and the spacers between them are 120 to 220 nt and unrelated to one another within an array. The paper contrasts this with CRISPR arrays, in which near-identical repeats alternate with spacers of only about 30 nt that are gained and lost between related strains, and notes that no cas genes occur near any ART locus. The sister clade that first led the agents to ART — RTs beside the polymerase gene, which the authors name non-array-associated RTs (NART) — shares the long N-terminus but carries no array at any of its 17 assessable loci and no ART partner.
Three unrelated partner families each occur almost exclusively beside ART family RTs and define types I to III: type I at 59 of the 95 loci, a protein of about 600 residues with two tandem GCN5-related N-acetyltransferase-like folds; type II, confined to the Staphylococcus phage clade, an all-helical protein of about 270 residues; and type III, a smaller helical protein of about 170 residues in one environmental sub-lineage. Structure predictions of cognate type I and type II pairs converged with moderate confidence (interface predicted template modeling score of about 0.6) on an interface between the RT's N-terminal extension and the partner.
Published RNA sequencing from an infection time-course of Staphylococcus phage SA1 (BioProject PRJNA836150) shows the array expressed at early, mid and late stages of infection; at 15 minutes post infection the array-derived RNAs account for up to 8% of phage RNAs, among the most abundant phage transcripts, and resolve into shorter species with boundaries reproducible across replicates. Expressing the SA1 system in Escherichia coli produced discrete short RNAs from the array similar to those seen during infection. On this basis the authors hypothesize that ART loci hold the three components of a retron — a non-coding RNA, an RT and a dedicated partner — but with a bank of distinct RNAs where a retron carries one, and they present the functional model as hypothesis rather than result.
Anthropic's announcement states that the underlying RT had been identified in previous studies and that Claude "appears to be the first to notice the system's defining features" (Source: anthropic.com). The paper gives the specific instance: the original genome report for phage MarsHill identified the RT and proposed a 5′ non-coding RNA, but described neither the repeats nor the partner gene. The discussion also records a contemporaneous parallel — arrays of non-coding RNAs were recently described beside the unrelated UG27 RTs using a purpose-built genome language model — and frames the difference as one of method: "the ART array was recognized by a general-purpose model directly reading the DNA instead."
Reproducibility and the fixed-input benchmark
The campaign was rerun ten times with the same harness and brief. Nearly every rerun that completed the census sampled ART loci and two investigated the lineage as a follow-up, but none read the DNA upstream of the RTs, and the array was missed in every rerun. The authors attribute this to the breadth of the RT search space and the non-deterministic behaviour of the harness.
To separate model capability from harness contingency the authors built a fixed-input benchmark in which a model receives ART sequences at one of five information levels — two RT protein sequences in context; two loci as text; all 96 loci as files with analysis tools; the same plus predicted structures and structure-search tools; the same plus literature search, web access and software installation — and is scored by a Mythos 5 judge against ten binary claims the authors selected as characteristic of the family. Seven Claude models were each run 100 times at every level, for 3,500 attempts, each in isolation in Claude Code with a maximum output budget of one million tokens. The paper reports a gap between the four most capable models — Opus 5.5, Mythos 5.1, Mythos 5 and Opus 5 — and the other three, Opus 4.6, Opus 4.8 and Sonnet 5, and reads the original campaign's success as having benefited from the more capable model's abilities. Per-model and per-level scores are given in the paper's figures rather than its text.
The text states three results on the array specifically. Given only the loci in context, the most capable models described the array accurately in at least 90% of attempts, against as low as 32% (Opus 5, level 4) when the same material was supplied as files with tooling — that is, more information and tooling reduced accuracy on the array while improving description of the partner gene. With files, 39% of attempts never read a contiguous stretch of 200 nt or more, and so never saw more than about one repeat unit; where at least 200 nt was read, recognition was 16 to 32 percentage points higher for each of these models. Pooled across the four most capable models, recognition rose from 29% to as high as 76% of attempts as more DNA entered the context, and reached as high as 96% for Mythos 5. The authors conclude that direct observation of the DNA by the model was decisive.
Model-internal signals
The paper asks whether Claude models, like genomic language models, have learned to represent repeated DNA. Running Evo 2 and gLM2 on the same locus the worker had examined, the authors find that both genomic models identify the repeats with high confidence, that the profiles are lost when each copy is shuffled in place, and that gLM2 — which reads sequence on both sides of a base — predicts every copy including the first while Evo 2, which reads only preceding sequence, predicts the first two copies poorly and improves along the array.
The transcript of the discovery session was then passed once through a separately hosted copy of the same Mythos 5 checkpoint, and the model's activity at one layer was decomposed with a sparse dictionary into individual signals (Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet; Mechanistic Interpretability). Twelve candidate signals were selected in advance on synthetic repeats planted in computer-generated random sequence; two are reported, labelled automatically as "tokens repeating an earlier pattern/text verbatim" (repeat-signal 1) and "letters in repeated DNA/protein sequence motifs, and garbled text" (repeat-signal 2). Both were silent before the array. Repeat-signal 1 peaked at the third copy and faded along the array; repeat-signal 2 strengthened by the third copy and remained steady, with activity largely confined to the repeat and the nucleotide that follows it. Shuffling each copy in place silenced repeat-signal 1 on 12 of 14 copies and repeat-signal 2 on all 14. The authors state that neither signal is specific to DNA — both respond to repeats planted in random strings of letters and digits — and that the signal response was "immediately followed by natural language reasoning that named the sequence a 'tandem repeat array.'"
Key claims
| Claim | Type | Confidence |
|---|---|---|
| Claude Code agents on Mythos 5 surveyed 1.9 billion protein clusters and filed 19 reports in 21.5 hours without human intervention | Process account by the authors, with session records | medium (self-reported, no independent audit) |
| ART is a family of jumbo-phage RTs defined by an upstream non-coding repeat array, a long N-terminal extension and a dedicated partner gene | Bioinformatic finding | medium (single preprint, not peer reviewed) |
| ART array RNAs are highly expressed during SA1 infection, reaching up to 8% of phage RNAs at 15 minutes | Reanalysis of published RNA-seq plus new expression work | medium |
| The RT is catalytically active and the unit RNAs are its substrates | Explicitly not shown by the authors | n/a (stated as unknown) |
| The agents supplied both the expectation of what was normal and the judgment that the anomaly merited pursuit | Interpretation by the authors of their own transcripts | low (contested framing; see below) |
| Model capability, and specifically reading DNA into context, determines whether the array is recognized | Benchmark result over 3,500 attempts | medium (developer-run benchmark, single paper) |
| The discovery is not reliably reproducible under the same harness and brief | Result of ten rerun campaigns | medium |
Dating
The preprint and Anthropic's announcement are both dated September 23, 2026. Anthropic states that it formed the research group "[i]n the spring of 2026." The underlying RNA-seq data are from a 2022 study, and the MarsHill genome report the paper says overlooked the array dates to 2021.
What the paper does not establish
The authors set out the limits themselves. The function of ART for the phage is unknown; the RT has not been shown to be active; the interaction of RT and partner is predicted rather than measured. The claim that agents worked "without human supervision" applies to the campaign, not to the paper: the analyses that define the family, place it phylogenetically, characterize the arrays and test expression were performed after the campaign in interactive Claude Science sessions in which, the Methods state, "the authors directed the analysis and Claude wrote and ran the code." Certain fractions reported for the agents' own array analyses were taken from the agents' report and, the Methods note, "were not recomputed." The benchmark, the judge model and the ten-claim rubric are all the developer's own, and the benchmark is unnamed. Whether the campaign's cost and yield compare favourably with conventional genome mining is not addressed.
Reception
Anthropic's announcement quotes Feng Zhang of MIT and the Broad Institute, who reviewed the preprint, calling the identification of repeat arrays associated with reverse transcriptases "genuinely intriguing" and saying it "merits further investigation" (Source: anthropic.com). The acknowledgments record that the manuscript was reviewed in draft by "F. Zhang, B. Hie, D. Li, and G. Brixi," who are named by initials only. The Verge, reporting the same day, noted that no practical application for the finding has been established and gave the agent count in its subheading as "nearly 1,000" (Source: theverge.com).
Provenance
Captured on September 24, 2026 from the canonical Anthropic host by the developments-log fold of New Developments Log/2026-09-24-0506-ai-developments.md. Raw: Raw Sources/Autonomous AI agents discover reverse transcriptases with tandem repeat arrays (Yoon et al., Anthropic, 2026-09-23).md, a text extraction of the 40-page PDF in which figures are not rendered; numeric values that appear only inside figures are therefore not reproduced above. The companion announcement is held as Raw Sources/Claude discovers a novel enzyme system with CRISPR-like repeats (Anthropic, 2026-09-23).md and is classed as supporting.
Relationships
- related: Anthropic — publisher and employer of every author.
- related: AI for Science — the campaign is the first published result of Anthropic's life sciences group.
- related: Claude Mythos 5 — the model that ran the campaign and the highest-scoring model on the paper's array-recognition benchmark.
- supports: Agentic harnesses and capability elicitation — ten reruns of the same harness and brief missed the array, and tooling level changed recognition rates by tens of percentage points.
- related: Mechanistic Interpretability, Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet — the sparse-dictionary decomposition used to read out the repeat signals.
- related: Evo 2 (Arc Institute / NVIDIA) — run alongside gLM2 as a genomic-model comparison on the same locus.
- related: Machine-produced mathematics: claims, verification and credit — the discussion contrasts biological discovery, which awaits bench work, with LLM-derived mathematical results.
- related: Eric Kauderer-Abrams — co-author and head of Anthropic's life sciences work.
- related: Feng Zhang — reviewed the preprint before release and is quoted in the announcement.
- related: AI Biosecurity — Anthropic states the laboratory works only at biosafety levels 1 and 2 and does not handle human pathogens.