Notes from Việt Nam on law, technology, languages, history, culture and travel.

The Agent That Read the DNA: What Anthropic's "Claude Discovered an Enzyme System" Claim Actually Shows

On 23 September, Anthropic’s new Life Sciences research group, which the company says it formed in spring 2026, announced that Claude had autonomously discovered a novel enzyme system. The system is a previously undescribed family of bacteriophage reverse transcriptases that sit next to long arrays of repeated DNA, a layout reminiscent of CRISPR. The team named it ART (array-associated reverse transcriptases). The claim spread quickly across outlets including Unite.AI, The Next Web and Investing.com, usually summarised as “950 agents, 21 hours, a CRISPR-like discovery”.

The accompanying technical report, “Autonomous AI agents discover reverse transcriptases with tandem repeat arrays” (Yoon, Athukoralage, Ameisen, Kauderer-Abrams, Perry, Durrant), is worth reading in full. The discovery is real, but it is narrower than the headline. The details that matter most to engineers building agent systems are mostly in the methods and the negative results.

What happened

The researchers gave a multi-agent harness a short research brief: find new reverse transcriptase (RT) systems, meaning enzymes that copy RNA into DNA, by looking for new partner genes that sit beside them in genomes. According to the preprint, every agent was a Claude Code instance running Claude Mythos 5. The campaign then ran with no human intervention:

  • The agents built 52 profile HMMs (statistical sequence models) of their own and searched 1.94 billion metagenomic protein clusters.
  • They recovered ~198,000 RT clusters, sorted them into nine classes, sampled ~11,000 loci and scored 3,564 candidate partner families.
  • 16 families passed the agents’ own selection criteria. A 17th was promoted by a follow-up task the agents opened themselves.
  • In total the run covered 119 tasks and 949 agent sessions: 77 agent-hours and 215.6 million tokens over 21.5 hours of wall-clock time. Press coverage rounded this to “950 agents” and “210 million tokens”.

The output was 19 reports for human review. Most candidates did not survive: of the 17 partner families, 14 were rejected as annotation artifacts, parts of known systems, or unrelated neighbouring genes. Three held up as genuinely new partner associations, and three new RT lineages were flagged. One of those lineages was ART.

The technical substance: an agent that looked at raw sequence

What makes the story interesting is how ART was found. The brief asked about partner proteins, but ART is defined by non-coding DNA. The session transcripts show a chain of agent decisions:

  1. A worker agent rejected a spurious gene association but still queued a follow-up on an RT family it found odd: a retron-like RT inside a jumbo phage.
  2. A supervisor agent reasoned that retron RTs use an upstream non-coding RNA, so it asked the next worker to inspect the 5′ flanking DNA.
  3. That worker pulled ~2,900 nucleotides of raw sequence into its context. It then wrote, in effect, “I can see by eye a tandem repeat array”, and pointed to a ~16–17 bp motif repeating with 100–200 nt spacers.
  4. It questioned its own novelty claim (“Actually wait — reconsider…”), wrote a repeat-counting script, compared the array against known elements including CRISPR, and ran a literature “kill test” before filing its report.

Humans then followed up, partly in interactive Claude Science sessions. They found 95 ART-family RT clusters, 28 with a detectable array. The arrays span 0.3–4.1 kb and hold 3–21 copies of a 15–49 nt palindromic repeat. The RT has an unusually long N-terminal extension of ~180 residues, where comparable RTs have ~50 or fewer. The partner proteins fall into three unrelated families. The arrays are not CRISPR: the spacers are 120–220 nt instead of ~30, the spacers are conserved between related phages instead of gained and lost, and no cas genes appear nearby.

A published RNA-seq time course of Staphylococcus phage SA1 showed the array transcripts making up up to 8% of phage RNA at 15 minutes post-infection. The team also expressed the SA1 locus on plasmids in E. coli and saw the array processed into discrete short RNAs. Their working hypothesis is a retron-like system that works with a bank of RNAs instead of a single one.

The part developers should read: it didn’t reproduce

The team ran the same campaign ten more times. None of the reruns rediscovered the array. Several reruns sampled ART loci, and two investigated the lineage, but no worker read the upstream DNA. The authors attribute this to the large search space and the non-deterministic harness.

So they built fixed-input benchmarks, and those produced the most practically useful finding in the paper. When models were given the ART loci directly in context, the four strongest (Opus 5.5, Mythos 5.1, Mythos 5, Opus 5) described the array correctly in ≥90% of attempts. When the models were given files and tools, performance dropped, to as low as 32% for one configuration. The reason was simple: in 39% of tool-using attempts the model never read a contiguous stretch of ≥200 nt, so it never saw more than one repeat. More tooling made the agents worse at noticing what was in front of them.

Anyone who has watched a coding agent grep around a file it should simply have opened will recognise this failure mode. The paper also includes an interpretability section. It identifies two internal Mythos 5 “repeat signals” that switch on as the array is read and go silent when each copy is shuffled. The authors present these as mechanistic evidence for what they call “genomic vision”.

Why it matters

Genome mining is bottlenecked on expert curation. Pipelines only find the novelty you predefine, and humans can’t read billions of sequences. The paper’s real claim is that a general-purpose model, reading raw data, can supply its own expectation and flag the anomaly, a step earlier automation left to humans. That pattern (agents doing broad, cheap, fallible screening and humans doing validation) carries over well beyond biology, to log analysis, security triage, and large-codebase audits.

Reasons for caution

  • Function unknown. The authors state plainly that they “have not shown that the RT is active or that the unit RNAs are its substrates”. The “cutting, copying, pasting DNA” comparison in the announcement describes other systems with similar features, not ART.
  • Self-reported, not peer-reviewed. Every step (models, harness, judge model for the benchmarks, interpretability) is Anthropic’s own. The one outside voice, CRISPR pioneer Feng Zhang, calls the find “genuinely intriguing” and says it “merits further investigation”. That is an endorsement of interest, not a validation.
  • Not first of its kind. CRISPR-like non-coding arrays beside the unrelated UG27 RTs were recently found with a bespoke genome language model, as the paper itself notes.
  • One hit, 1-in-11 reproducibility. A result that surfaced in one run out of eleven is a strong demo and a weak method. The discovery may say as much about luck and harness design as about model capability.

It is a genuine, well-documented anomaly-detection result, and a candid case study in how fragile agent behaviour still is. Whether ART turns out to be a useful biotech tool is a question for the wet lab.

Sources