Claude agents flagged a new phage enzyme family. What it does is still unknown.

AI FOR SCIENCE · GENOME MINING

An agent was working through 2,900 letters of raw phage DNA when it stopped and wrote: “I can see by eye a tandem repeat array … that's a CRISPR-like or msDNA-like repeat array?!” Anthropic's preprint quotes the line, and it marks the start of a new family of phage enzymes. The company calls it ART, for array-associated reverse transcriptases. Nobody yet knows what ART does. The better story is how an agent system found something its own brief never asked for.

QUICK ANSWER

A real anomaly, found by agents; the biology is still open

The evidence supports a narrow claim: an autonomous harness surveyed reverse transcriptases at scale and flagged an unannotated repeat array that human curators had missed. It stops short of a new gene-editing tool.

What the agents were asked to do

Reverse transcriptases (RTs) copy RNA into DNA. In bacteria they often sit beside a partner protein and a non-coding RNA and work as antiphage defenses. Retrons are the best-known case. The brief told the agents to look for new RT systems through new partner-gene associations. It never mentioned repeat arrays.

The funnel, as reported in the preprint:

Stage	Output
Build search profiles	52 RT profile HMMs, built by the agents
Search	1.94 billion metagenomic protein clusters
Curate and classify	198,290 RT clusters in 9 classes
Neighbor census	~11,000 loci sampled, 3,564 candidate partner families scored
Deep dives	16 families plus 1 promoted by a follow-up task
Outcome	3 new partner associations, 14 set aside, 3 new RT lineages, 19 reports

The announcement rounds these to about 950 agents, 210 million tokens, 21 hours, 200,000 RTs, 3,500 candidates and 20 reports. Where the two disagree, the preprint has the more exact count.

Inside the harness

Human research brief→Stages split into tasks→Worker agent plans and runs→Supervisor agent reviews, opens follow-ups→Reports ranked for human review

Each task pairs two Claude Code agents running Claude Mythos 5. A worker proposes a plan and carries it out. A supervisor reviews the plan and the results, and can open new tasks when something odd turns up. Everything lands in a shared record, so agents build on each other's work. Of 119 tasks, 98 were opened by agents, not seeded by the brief. The 19 final reports were ranked in a pairwise tournament before humans picked what to test.

The discovery came from a follow-up nobody planned: an agent pulled flanking DNA to check a retron-like signal, then read the sequence itself.

What ART looks like

ART sits mainly in jumbo phages and has three parts: an RT, a dedicated partner gene and a long array of units about 200 nucleotides long. Each unit carries a short repeat core, such as CATGTGTATCGCATGT, followed by a unique spacer of 120 to 220 nucleotides. The first locus the agent flagged had 14 copies of a 16-nucleotide repeat.

The resemblance to CRISPR is in the layout, not the details. CRISPR repeats are near-identical, with spacers of about 30 nucleotides that are gained and lost between strains. ART units are longer, and their order is conserved between related phages such as MarsHill and SA1. The three partner families share no detectable sequence or structural similarity, which hints that the RT was paired with unrelated partners more than once.

The lab evidence so far

• Expression during infection: a Claude Science session found and reanalyzed RNA reads from a published time course of Staphylococcus phage SA1 infection. The arrays were highly expressed at early, middle and late stages.
• Expression without the phage: the team expressed the SA1 ART system on plasmids in E. coli and ran small-RNA sequencing. The array produced distinct short RNAs, similar to those seen during infection.
• Not yet shown: in the authors' words, they “have not shown that the RT is active or that the unit RNAs are its substrates.” Whether the RT and partner interact, and what ART does for the phage, is unknown.

All lab work was done by human scientists at biosafety levels 1 and 2, with no pathogens that infect humans.

Earlier work and credit

The RT itself was already known. The original MarsHill genome report identified it and proposed a non-coding RNA upstream, but described neither the repeats nor the partner gene. RNA arrays next to a different RT family, UG27, were recently found with a purpose-built genome language model, so this architecture appears to have evolved more than once. TechCrunch reports that Dario Amodei acknowledged a Stanford team had found a system “in some ways similar.” The same TechCrunch article says the system “can perform operations like cutting, copying, and pasting DNA.” Anthropic's post says only that related systems do, and that ART's function is unknown.

Feng Zhang, who reviewed the preprint, called the RT-associated repeat arrays “genuinely intriguing” and said they merit more study (quoted in Anthropic's post).

Try the idea: a toy repeat finder

The worker found the array with a repeat-counting script it wrote after spotting the pattern by eye. The sketch below shows the basic idea on synthetic DNA. It is a teaching example. It is not the preprint's code and not a genomics tool. Real arrays have mismatches, which exact k-mer matching misses.

from collections import defaultdict

def tandem_candidates(seq, k=16, min_copies=4, min_gap=80, max_gap=260):
    """Flag k-mers that recur at spacer-like distances."""
    seq = seq.upper()
    hits = defaultdict(list)
    for i in range(len(seq) - k + 1):
        hits[seq[i:i + k]].append(i)
    found = []
    for kmer, pos in hits.items():
        gaps = [b - a for a, b in zip(pos, pos[1:])]
        spaced = [g for g in gaps if min_gap <= g <= max_gap]
        if len(pos) >= min_copies and len(spaced) >= min_copies - 1:
            found.append((kmer, len(pos), min(spaced), max(spaced)))
    return sorted(found, key=lambda x: -x[1])

# Synthetic test: 6 copies of the repeat core, random 110-170 nt spacers
# Output from our run:
# CATGTGTATCGCATGT  copies=6  distance=130-150 nt (start to start)

Why the method matters more than the enzyme

Genome-mining pipelines decide in advance what counts as new, and anything outside those features is invisible. The ART array fell outside the brief, so a pipeline built from that brief would have skipped it. What worked here was a structure that let agents chase their own observations and made a second agent approve those detours. Most of the 17 candidate families were rejected, and that is a good sign: the supervisors look like filters, not rubber stamps.

The bottleneck moves rather than disappears. Nineteen reports is a readable pile. A campaign ten times larger would produce hundreds, and the preprint admits the team is still studying what separates reports worth testing from the rest. Human judgment stays the scarce resource. It has just moved from the middle of the funnel to the end.

The preprint also claims two internal Mythos 5 signals respond to repeated DNA in a similar way to models trained only on biological sequences. That is interesting, but it is the company's analysis of its own model, and outside groups cannot check it yet.

Limits to keep in view

• This is a company preprint announced by the company, not a peer-reviewed paper.
• “Autonomous” covers the computational campaign. The brief, curation choices and all experiments were human.
• The CRISPR comparison is about layout. No editing, cutting or targeting activity has been shown.
• The token count shows the scale of compute. The paper does not give a dollar cost.
• Biology agents raise dual-use questions. Anthropic says the lab works only at BSL-1 and BSL-2 and that humans run every experiment.

What to watch next

1. Biochemical evidence that the ART RT is active, and on which RNA.
2. Whether the partner protein binds the RT and the unit RNAs.
3. Independent labs finding ART loci in other phage genomes, or reproducing the expression results.
4. Whether other groups can run a similar harness on another protein family and get a comparable hit rate.

Verdict

ART is a real new entry in the growing list of prokaryotic RT systems with CRISPR-like arrays, and its function is open. The stronger result is about method: a supervised multi-agent harness read raw sequence, noticed something outside its brief and backed it with enough evidence for humans to test. Treat the CRISPR framing as a comparison of shapes until the biochemistry arrives.

FAQ

Did Claude discover a new CRISPR?

No. It flagged a phage enzyme system with a CRISPR-like repeat layout, and no gene-editing activity has been shown.

How autonomous was it?

The 21.5-hour computational campaign ran without human steps. Humans wrote the brief, reviewed the reports and did all lab work.

What is known about ART's function?

Only that its arrays are transcribed into distinct short RNAs. Its enzyme activity, substrates and biological role are all unknown.

Anthropic announcement (Sep 23, 2026) · ART preprint (PDF) · TechCrunch · The Verge