Autonomous AI Agents Discover Reverse Transcriptases with Tandem Repeat Arrays

The first published result of Anthropic's new life-sciences research group and laboratory, which Anthropic introduced alongside this preprint: a genome-mining campaign run by Claude agents that found a previously undescribed family of phage enzymes. The research brief asked for novel reverse transcriptase (RT) systems defined by new partner genes. An agentic harness turned it into stages and tasks, each carried out by a Claude Code worker whose plan and results a supervisor agent reviewed; supervisors could open follow-up tasks, and every plan, result and review went to a shared record. Running on Claude Mythos 5 without human intervention, the agents built their own profile HMMs, searched 1.9 billion metagenomic protein clusters, recovered about 200,000 RT clusters in nine classes and scored 3,564 protein families found near RTs as candidate partners. The campaign ran 119 tasks (98 opened by the agents themselves) in 949 agent sessions, using 77 agent-hours and 215.6 million tokens over 21.5 hours of wall-clock time, and ended with 19 reports. Of 17 candidate partner families, three were confirmed as previously unreported RT associations, and three new RT lineages were flagged.

The headline find came from outside the brief. A worker reading raw DNA beside a jumbo-phage RT noticed an unannotated tandem repeat array and wrote: "I can see by eye a tandem repeat array … that's a CRISPR-like … repeat array?!" The authors named the family array-associated reverse transcriptases (ART): an RT, a dedicated partner gene and an array of units of about 200 nucleotides, with no cas genes nearby. A Claude Science session then retrieved public RNA-seq data from an infection time course of Staphylococcus phage SA1, in which the arrays are highly expressed and resolve into discrete short RNAs; small-RNA sequencing of the SA1 system expressed in E. coli gave similar RNAs. The authors suggest a retron-like system in which one enzyme works with a bank of distinct RNAs, but whether the RT is active and what the system does for the phage are still unknown. Feng Zhang of MIT and the Broad Institute, commenting on the preprint, called the RT-associated RNA-repeat arrays "genuinely intriguing".

The discovery did not reproduce: ten reruns of the campaign with the same harness and brief sampled ART loci and twice followed up the lineage, but no agent read the upstream DNA, and the array was missed every time. In fixed-input benchmarks judged against ten curated ART features, four models (Opus 5.5, Mythos 5.1, Mythos 5 and Opus 5) clearly outperformed Opus 4.6, Opus 4.8 and Sonnet 5. Given the loci in context, the strongest models described the array in at least 90% of attempts, but with files and tools the rate fell as low as 32%. In 39% of those file-based attempts the model never read 200 or more contiguous nucleotides (about one repeat unit), and attempts that did recognized the array 16 to 32 percentage points more often. An interpretability analysis of the original session found Mythos 5 internal signals that respond to the repeated DNA (silenced when each copy is shuffled) just before the model called it a "tandem repeat array". All six authors are at Anthropic, with Nicholas T. Perry and Matthew G. Durrant corresponding; the 40-page preprint is a PDF on Anthropic's site.