Anthropic's Claude Agents Find a New CRISPR-Like Phage Enzyme Family in 21.5 Hours

TL;DR

Anthropic says a fleet of Claude agents found a previously undescribed family of phage enzymes, which it calls array-associated reverse transcriptases (ART). The campaign ran 949 agent sessions on Claude Mythos 5 over 21.5 hours of wall-clock time and 215.6 million tokens, surveying 1.9 billion metagenomic protein clusters with no human intervention. The find came from one agent reading raw DNA next to a phage enzyme and noticing an unannotated repeat array. Lab work confirms the array is transcribed into distinct short RNAs, but nobody yet knows what the system does, and ten reruns of the identical campaign missed it.

What was found

The subject is the reverse transcriptase (RT), the enzyme family that copies RNA into DNA. In bacteria, RTs often sit next to a partner protein and a non-coding RNA, the setup behind retrons, which bacteria use to defend against viruses.

ART is a new lineage of RTs found mostly in jumbo phages, the oversized viruses that infect bacteria. According to the technical report, each ART locus carries three parts:

• An array upstream of the RT: 3 to 21 copies of a short repeat (15 to 49 nucleotides) at near-constant spacing, with unrelated 120 to 220 nucleotide spacers between them.
• The RT itself, with an unusually long N-terminus of about 180 residues where other RTs carry about 50 or fewer.
• A dedicated partner gene directly downstream, in three unrelated types that share no detectable sequence or structural similarity.

The array is why the press is saying "CRISPR-like." A CRISPR array also alternates repeats with spacers. The resemblance stops there: ART has no Cas genes, its spacers are conserved between related phages instead of freshly acquired, and the report is explicit that ART "has not been shown to function like CRISPR."

The underlying RT was not new either. It had already been identified in a jumbo phage in earlier work. What the agents added was the surrounding system: the array and the partner, which together define a family nobody had described.

How the agents worked

The harness is ordinary engineering, which is the interesting part. Every agent was a Claude Code instance configured with Mythos 5, running in a sandbox with 60 CPU cores, 192 GiB of memory, no GPU, and up to 58 concurrent sessions. Structure predictions went to a GPU queue (19 jobs total).

A launch agent turned the research brief into five chained stages: input assembly, database sweep, RT classification, neighborhood census, and deep dives. Each task went to a worker that wrote and executed a plan; a supervisor agent accepted the result or sent it back. A curator agent logged findings into a shared knowledge base that later workers saw in their prompts, and an editor agent reviewed reports before filing.

The tooling is the standard bioinformatics kit: HMMER, MMseqs2, BLAST+, geNomad, and structure prediction with ESMFold or ColabFold. The retained logs show 7,578 shell commands, 696 database queries, 131 literature searches, and 61 web requests. Agents also had a library of about 140 written skills, method guides for things like testing novelty claims.

The funnel

The database held predicted proteins from roughly 11 million biosamples, clustered into 1.94 billion protein clusters. The agents built their own search profiles, recovered about 200,000 RT clusters, sorted them into nine classes, sampled about 11,000 RT loci, and scored 3,564 protein families that kept showing up next to RTs as candidate partners.

Sixteen families passed the agents' own selection criteria, and a seventeenth was promoted from a follow-up task. The agents were also honest about their hit rate: only three of the 17 held up as previously unreported RT associations. The other 14 were rejected as annotation artifacts, parts of known systems, or neighbors that just happened to be nearby. The campaign closed with 19 reports, three of them for new RT lineages that workers flagged on their own.

The moment that mattered

The brief asked for new partner genes. ART showed up somewhere the brief never pointed: the non-coding DNA upstream of the enzyme. The team went back through the session transcripts to see how. A worker had loaded the raw DNA flanking a jumbo-phage RT into its context, read it, and wrote that it looked like "a CRISPR-like repeat array." A scan then found one locus with 14 copies of a 16-nucleotide repeat, and the agents filed a report.

This is the hard part to picture, so here is the analogy. Imagine being handed a song's lyrics as one unbroken string of letters, with no line breaks, and noticing the chorus coming back every 200 characters. A pipeline only finds the patterns it was told to count. The model noticed a pattern it was never asked about, because it actually read the sequence.

Anthropic's interpretability team backed that up. Passing the discovery transcript through Mythos 5 again, they found two internal signals that fire as the model reads the repeat copies. Shuffling each copy in place silenced one signal on 12 of 14 copies and the other on all 14.

Reproducibility is the weak spot

Anthropic ran the same campaign ten more times. Nearly every run sampled ART loci and two investigated the lineage as a follow-up, but none read the upstream DNA, and the array was missed in every rerun. Science, it turns out, also has flaky tests.

So the team built fixed benchmarks: give a model ART sequences, either directly in context or as files with tools, and have a judge model score its report against ten curated features. Seven Claude models split cleanly into two groups. Opus 5.5, Mythos 5.1, Mythos 5, and Opus 5 formed the capable group; Opus 4.6, Opus 4.8, and Sonnet 5 fell behind.

The counterintuitive result: more tooling made the array harder to find. With the loci in context, the capable models described the array in at least 90% of attempts. With files and tools, that fell to as low as 32% (Opus 5). In 39% of file-based attempts, the model never read a contiguous stretch of 200 nucleotides, so it never saw more than about one repeat unit. As models read more DNA, recognition climbed from 29% to 76% pooled, and to 96% for Mythos 5.

If you build agent harnesses, that is the transferable lesson. Tools that summarize data for the model can hide exactly the anomaly you hoped it would catch. Sometimes the right move is to make the agent look at the raw bytes.

What the lab showed, and what it did not

Humans took over after the reports. In interactive Claude Science sessions, the team found 95 distinct ART RT clusters, 28 of them with a detectable array. At the bench, they checked RNA data from Staphylococcus phage SA1: at 15 minutes after infection, array-derived RNAs made up as much as 8% of phage RNAs, among the most abundant phage transcripts. Expressing the SA1 system from plasmids in E. coli produced the same discrete short RNAs.

The hypothesis is a retron-like system with a bank of RNAs instead of one, so a single enzyme and partner could form several complexes that respond to different triggers. The report is blunt about the gap: the team has not shown the RT is active, that the RNAs are its substrates, or that the RT and partner interact. What ART does for the phage is unknown.

CRISPR pioneer Feng Zhang, of MIT and the Broad Institute, called the identification of the RNA-repeat arrays "genuinely intriguing" and said it "merits further investigation," per The Next Web. The report also notes the architecture is not unique: arrays beside unrelated UG27 RTs were recently found with a purpose-built genome language model. ART's twist is that a general-purpose LLM spotted it by reading, not a model trained only on DNA like Evo 2.

Why builders should care

Strip out the biology and this is a case study in long-running autonomous agents doing useful work: 119 tasks, 77 agent-hours, a supervisor on every task, a shared memory, and 14 of 17 candidates correctly thrown out by the agents themselves. The pattern (worker plus reviewer, curated memory, agents allowed to open their own follow-ups) is reproducible with public tooling.

The caveats are just as useful. The discovery depended on a frontier model that most organizations cannot buy, on one agent's choice to read raw input, and on luck that ten reruns did not repeat. Treat it as proof that the ceiling exists, not as a pipeline you can run tomorrow and expect a new enzyme family by Friday.

Key Takeaways

• Anthropic's agents surveyed 1.94 billion protein clusters in 21.5 hours using 949 Claude Code sessions on Mythos 5 and 215.6 million tokens, with no human intervention.
• ART is a new jumbo-phage RT family with a CRISPR-like repeat array and a dedicated partner gene; the RT was known, the system was not.
• Its function is unknown: the lab shows abundant short RNAs from the array, but not that the enzyme is active.
• Ten identical reruns missed the array, so the discovery was real but not reproducible on demand.
• Benchmarks show giving models more tools made them worse at spotting the array, because they stopped reading the raw DNA.
• The harness pattern (worker, supervisor, curator, editor, shared memory) is the part you can copy today.

Sources: Anthropic announcement, Anthropic technical report (PDF), The Next Web, Interesting Engineering