Claude agents spotted a new family of phage enzymes. What it does is still unknown.

On 23 September 2026 Anthropic said a fleet of Claude agents, searching about 1.9 billion protein clusters without human intervention, had turned up a family of reverse transcriptases nobody had described: array-associated RTs, or ART, found mostly in jumbo phages. The 40-page preprint linked from the announcement closes its own account on the limit: "we have not shown that the RT is active or that the unit RNAs are its substrates. Whether the RT and its partner interact, and what the system does for the phage, are currently unknown."

Carry that sentence into any conversation where ART is sold as the next CRISPR. The announcement leans that way itself, saying the system "has a set of characteristics that have only ever been found together in a handful of other systems, all of which are programmable and perform operations like cutting, copying, and pasting DNA." The preprint claims no application. Neither should anyone quoting it.

If you build or run agents on scientific data, the paper holds a sharper finding, and it is about how the agent looked rather than what it found.

The preprint's counts are not the announcement's

Every agent was Claude Code running Claude Mythos 5, up to 58 sessions at once in a sandbox with 60 CPU cores and no GPU. The agents built 52 profile HMMs, searched 1,939,242,578 protein clusters, and kept 198,290 RT clusters, which they sorted into nine classes. They sampled 10,983 loci around those RTs, scored 3,564 recurring neighbour families as candidate partners, promoted 16 and later a 17th, and confirmed three as new RT associations. Separately they flagged three new RT lineages. The campaign closed with 19 reports.

It ran as 119 tasks and 949 sessions: one launch session, 414 workers, 375 supervisors, 107 curators and 52 editors. That came to 76.9 agent-hours and 215.6 million tokens over 21.5 hours of wall-clock time. Of those tokens, 189.5 million were writes to the prompt cache, 14.9 million were output and 11.3 million were uncached input. Cache reads were not counted.

The announcement rounds, and not always in the same direction. It gives "roughly 950 agents" where the preprint counts 949 sessions across five roles, "210 million tokens", "21 hours" and "over 200,000 RTs". Its "3,500 new candidate systems" are, in the preprint, the 3,564 neighbouring protein families scored as possible partners. Its "20 most compelling candidates" match no count in the preprint, which has 17 candidate families and 19 reports.

An agent read the raw DNA and saw the repeats

The route to ART was sideways. The agents first picked this RT because it sat next to a jumbo-phage RNA polymerase gene. A worker agent judged that pairing spurious, then queued a follow-up anyway, because a retron-like RT inside a phage genome looked odd. Retron RTs work with a non-coding RNA encoded upstream, so the supervisor sent the next worker to check the upstream DNA. The closest relatives had long non-coding stretches there, a median of 940 bp. The worker pulled one 2,900-nucleotide flank into its context, returned without annotation, and wrote:

The L0050 (228,907 bp logan contig) flank is spectacular: I can see by eye a tandem repeat array

No repeat finder had run in that session, and neither the research brief nor the task brief mentioned repeats or arrays. A counting script the worker then wrote found 14 copies of a 16-nucleotide repeat, separated by unique spacers of 100 to 200 nucleotides. The agents checked the layout against CRISPR arrays and the known RT systems, searched the literature and filed a report. A tournament judged by Mythos 5 ranked it third of the 19.

The record before this matters. Phage MarsHill's 2021 genome report had already identified the RT and proposed a non-coding RNA upstream of it, yet described neither the repeats nor the partner gene. And another recent study found arrays of non-coding RNAs beside an unrelated RT family, UG27, using a genome language model, so the architecture is not unique to ART.

Humans directed everything after the report

"Our involvement was limited to the initial prompt and the lab work," the announcement says. That describes the search. The preprint's methods describe more: "All analyses that were performed after the campaign were carried out in interactive Claude Science sessions, in which the authors directed the analysis and Claude wrote and ran the code." Those sessions defined the family (95 members, 28 with a detectable array upstream of the RT), placed it in a clade next to the retrons, found three unrelated partner-gene types, and reanalysed a published 2022 infection time course of Staphylococcus phage SA1. In that dataset the array RNAs make up as much as 8% of phage RNAs 15 minutes after infection.

The bench work so far is one experiment. The team cloned parts of the SA1 ART locus onto two plasmids in E. coli, grew one culture per construct, and ran small-RNA sequencing. The array came out as discrete short RNAs, much like those seen during infection. The authors' working model is a retron-like system carrying a bank of distinct RNAs where a retron carries one. Their own figure draws its steps with dashed arrows, labelled hypothesised.

Ten reruns missed it

Anthropic launched the same campaign, same harness and brief, ten more times. Nearly every run that finished the census sampled ART loci, and in two of them workers followed the lineage up. None read the DNA upstream of the RTs, and "the array was missed in every rerun."

So the team built a fixed test: seven Claude models, 100 attempts per model at each of five input levels, graded by a Mythos 5 judge on ten features the authors chose. Four models (Opus 5.5, Mythos 5.1, Mythos 5 and Opus 5) clearly beat the other three (Opus 4.6, Opus 4.8 and Sonnet 5). The telling result runs against intuition. Given the loci as text in the prompt, the four strongest described the array in at least 90% of attempts. Given the same loci as files with tools, the rate fell as low as 32%, for Opus 5 at level 4. In 39% of their file-based attempts the models never read a contiguous 200 nucleotides, about one repeat unit.

For anyone wiring agents to sequence databases, that is the result to keep. The discovery came from a model looking at raw data, and the tooled setups often skipped the looking.

The preprint also traces two internal Mythos 5 signals that fired on the repeats as the worker read them. The methods add that "neither is specific to DNA": both respond to repeats planted in random strings of letters and digits too.

Feng Zhang of MIT and the Broad Institute, quoted by Anthropic after reviewing the preprint, said: "The identification of RNA-repeat arrays associated with reverse transcriptases is genuinely intriguing and merits further investigation." As of 24 September 2026, that is the right size for it. A well-documented anomaly, one expression experiment, and a hypothesis.

Sources

Both retrieved 24 September 2026.

• Anthropic, "Claude discovers a novel enzyme system with CRISPR-like repeats," 23 September 2026 — anthropic.com/news/claude-discovers-novel-enzyme-system
• P. H. Yoon, J. S. Athukoralage, E. Ameisen, E. Kauderer-Abrams, N. T. Perry and M. G. Durrant (Anthropic), "Autonomous AI agents discover reverse transcriptases with tandem repeat arrays," 40-page PDF linked from the announcement. It prints no date; its PDF metadata records creation on 23 September 2026 — www-cdn.anthropic.com/22573675ada52a8ca8a97a1a4b4326b2f208a071.pdf