At 19:30 UTC on September 23, Anthropic's biology announcement had reached 560 points and 567 comments on Hacker News. The less flattering number sits deep in the technical report: when the researchers launched the same broad search ten more times, every rerun missed the DNA repeat array that defined the original find. That gap makes the work more useful to anyone building research agents. One autonomous run surfaced a real lead. The route to it was rare and expensive, with the outcome sensitive to what the model happened to read.
Anthropic says a swarm of Claude agents identified a previously undescribed system in bacteriophages, the viruses that infect bacteria. The company calls it array-associated reverse transcriptase, or ART. Human scientists then found evidence that its repeat array is transcribed into short RNAs. They have not established what ART does, whether its reverse transcriptase is active or whether those RNAs are its substrates, according to the team's preprint. The discovery is a well-supported candidate for more biology, rather than a new gene-editing tool ready to rival CRISPR.
A 21-hour search produced one unusual trail
The project began with a research brief asking Claude to look for novel reverse transcriptase systems through new partner-gene associations. Reverse transcriptases copy RNA into DNA. The agents assembled search models, surveyed 1.94 billion protein clusters and recovered about 198,000 reverse transcriptase clusters after filtering. They sampled roughly 11,000 genomic neighborhoods, scored 3,564 recurring protein families and promoted 17 families for deeper analysis, the technical report says.
The orchestration was substantial. The campaign comprised 119 tasks and 949 Claude Code sessions, with worker agents doing analyses and supervisor agents reviewing their plans and results. Curator and editor agents maintained a shared record. Up to 58 sessions could run at once in a sandbox with 60 CPU cores and 192 GiB of memory. The work consumed 76.9 agent-hours over 21.5 hours of elapsed time and 215.6 million counted tokens. That token total excludes reads from the prompt cache, according to the methods section.
The useful branch began with a mistake. An agent first connected an unusual reverse transcriptase to a nearby phage RNA-polymerase gene, then rejected that association as incidental. Its supervisor asked a follow-up worker to inspect the noncoding DNA upstream. When the worker loaded the raw sequence into context, it noticed a repeating motif separated by longer, varying stretches of DNA. The original assignment concerned protein-coding partner genes, so this repeat array sat outside the search feature the researchers had specified, as the paper's task trace shows.
That detour is the strongest part of Anthropic's autonomy claim. Human researchers wrote the initial brief and built the harness, but the agents proposed 98 follow-up tasks as observations accumulated. The ART branch emerged from one of them. Anthropic says no person intervened during the 21.5-hour campaign, though humans reviewed the resulting candidate and performed the physical experiments afterward. The company's announcement describes the lab work as entirely human-run.
ART resembles CRISPR in layout, not proven function
The name needs care. ART has a reverse transcriptase beside a noncoding repeat array and a dedicated partner gene. Researchers identified 95 related reverse transcriptase clusters. Of those, 28 had a detectable upstream array. Those arrays ranged from 0.3 to 4.1 kilobases and contained three to 21 repeat copies. The short repeats were separated by longer sequences that differed from one another, an arrangement that visually recalls a CRISPR array, according to the preprint's sequence analysis.
The differences matter just as much. ART's spacers are 120 to 220 nucleotides long, while the report describes CRISPR spacers as roughly 30 nucleotides. Related phages retained ART spacers in the same order, unlike the rapid spacer turnover expected in CRISPR systems. The team also found no nearby cas genes. The evidence supports a new noncoding repeat element associated with a reverse transcriptase. It does not show a CRISPR-Cas system, and the authors explicitly leave ART's function unresolved.
There is experimental evidence beyond a sequence pattern. Claude located public RNA-sequencing data from a Staphylococcus phage SA1 infection. In those data, RNA from the ART array appeared throughout the infection and accounted for as much as 8% of phage RNA at 15 minutes. Human researchers then placed the SA1 system into E. coli and again observed discrete short RNAs from the array, the report says. That result confirms expression and processing of the array under the tested conditions.
It stops short of a mechanism. The team has not shown that the reverse transcriptase performs a reaction, that the short RNAs guide it or that the partner protein binds it. Nor has it demonstrated defense activity, programmable targeting or a benefit to the phage. Anthropic's public post says work on the primary function is continuing, while the preprint lists those open tests directly. The CRISPR comparison is therefore a clue about architecture and a reason to investigate, not a statement of equivalent capability.
The reruns reveal an attention problem
Anthropic tested whether the original discovery would recur by launching ten more campaigns with the same harness and research brief. Nearly every completed run sampled ART loci during its census, and two opened follow-up investigations into the lineage. None of the agents read the relevant upstream DNA. All ten runs missed the repeat array, which the authors attribute to the size of the reverse-transcriptase search space and the harness's nondeterministic behavior, according to the reproducibility section.
A more controlled test produced a revealing contrast. When the researchers placed ART loci directly in the prompt, the four most capable Claude models described the array correctly in at least 90% of attempts. Performance could fall to 32% when the same models received files and analysis tools instead. Across file-based attempts, 39% never read a continuous stretch of 200 nucleotides, so the model never saw more than about one repeat unit. Recognition improved by 16 to 32 percentage points when a model did read at least 200 nucleotides, the benchmark results show.
For agent developers, this is the transferable result. Giving a model access to more data and tools did not ensure that the relevant evidence entered its working context. In this experiment, added tooling sometimes reduced repeat recognition because agents inspected less raw DNA. That reading follows from Anthropic's benchmark rather than proving a rule for every agent system, but it identifies a concrete failure mode: an agent can possess the right files and tools while failing to look at the decisive 200 characters.
The evaluation also stays inside Anthropic's stack. Mythos 5 powered the original campaign, ranked the 19 agent-written reports in pairwise comparisons and graded the fixed-input benchmark against ten features selected by the authors. The report says the ART write-up placed third, winning 32 of 36 comparisons. That is useful process evidence, but it is not an independent scientific review. The paper is an Anthropic preprint authored by six Anthropic researchers, and the company released it alongside its new life-sciences lab.
Discovery here means finding the next experiment
The phrase "Claude discovered" compresses several kinds of work. Claude agents searched existing sequence databases, wrote and ran analyses, rejected an early hypothesis and flagged the pattern. Human scientists decided the lead warranted lab time, expanded the family analysis and ran the experiments that verified the array's RNA output. The paper's accounting supports genuine autonomy during the computational campaign, alongside a clear human boundary at physical validation and scientific interpretation.
That division is already consequential. Metagenomic collections contain far more sequences than specialists can inspect one by one. An agent that notices an odd arrangement outside a prescribed search can help researchers choose where to spend scarce bench time. Yet the ten failed reruns show why a single successful trace cannot become a dependable discovery pipeline on its own. Context selection and repeatability still determine whether an interesting locus is seen or silently passed over.
The next evidence should come from the molecule, not another agent transcript. Watch for experiments showing that the ART reverse transcriptase is active, identifying what it copies, testing whether the short RNAs direct that activity and establishing what the system does during phage infection. Independent replication and peer review would also separate a durable biological finding from an intriguing result produced inside one company's model-and-evaluation loop. Until then, Anthropic has shown that agents can find a good experiment. The experiment that explains ART is still waiting.