Claude flagged a CRISPR-shaped enzyme system. Nobody knows what it does.

Anthropic said on Wednesday that a swarm of Claude agents, searching protein databases without a specific target in mind, flagged an enzyme system nobody had described before: a reverse transcriptase sitting beside an accessory gene and a long array of evenly spaced DNA repeats — the architectural signature that makes CRISPR work. The company calls the family ART, for array-associated reverse transcriptases. In the same post, it says it does not yet know what ART does.

The announcement doubles as the public debut of Anthropic's life sciences research group, a Bay Area operation founded in spring 2026 that pairs model work with a wet lab running at biosafety levels 1 and 2, the two lowest containment tiers. The group publishes a technical preprint alongside the post.

What the agents found

ART has three parts: a reverse transcriptase, an enzyme that copies RNA back into DNA; an adjacent partner gene; and a stretch of non-coding DNA made of repeating units at regular spacing. Repeat arrays of that kind are uncommon, and only a handful of described systems combine all three features. Every one of those that has been characterised is programmable — it cuts, copies or pastes DNA to order. That is why the layout is interesting. It is also why the finding should be read carefully: a structural resemblance is a hypothesis about function, not a demonstration of one.

The reverse transcriptase at the centre of it was not itself new. Anthropic is explicit that this RT, carried by a jumbo phage, appears in earlier studies. What Claude appears to have noticed first is the combination — the repeat array next to it, and the accessory protein of unknown function. Human scientists did the bench work that followed, expressing the protein in standard laboratory strains and characterising it biochemically and structurally. Those first experiments indicate the ART array is itself transcribed into a set of distinct short RNAs, which is the sort of behaviour a programmable system would show.

The run, in numbers

For anyone building with agents rather than with enzymes, the shape of the search is as informative as the biology. Anthropic puts the campaign at roughly 950 agents, 21 hours of searching, and 210 million tokens consumed. The preprint sets out the funnel behind those figures:

• 52 reverse-transcriptase profile HMMs used as the starting queries
• a sweep across about 1.94 billion protein clusters
• 198,290 reverse-transcriptase clusters retained after filtering, sorted into 9 classes
• a neighbour census yielding 3,564 partner families across 10,983 loci
• roughly 20 finalists compressed into human-readable reports for review

The structure is a role hierarchy, not a flat swarm. Workers run tasks and propose follow-ups; a supervisor approves, revises or merges them; an editor ranks results by pairwise tournament; a curator maintains the shared knowledge base so that 950 agents do not spend the night re-deriving each other's work. Ninety-eight of the campaign's 119 tasks were follow-ups the system opened for itself. Anthropic says this class of analysis would take an expert scientist weeks to months.

The moment of the find sits in an agent's own log, which the preprint reproduces:

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

From there the agent counted the repeats, measured their spacing, compared the layout against known reverse-transcriptase systems and searched the literature for any prior report of the pattern before filing a report for a human to read.

What is not established

Function, first and most importantly. Anthropic states that its work to determine what ARTs primarily do is ongoing. Nothing in the announcement claims the system edits genomes, and nothing shows it doing so.

Novelty is qualified too, partly by Anthropic itself. Dario Amodei acknowledged the work was built on the work of others, and that a Stanford team had previously found a system similar in some respects to the one Claude surfaced. TechCrunch notes that the wider research community still has to judge whether the result is genuinely novel or significant, and that AI-driven biology is not Anthropic's alone — Stanford, UCSF and Google have pursued comparable work.

The outside voice Anthropic put forward is Feng Zhang, one of the central figures in CRISPR's development, at MIT and the Broad Institute:

This is an exciting example of how AI agents can contribute to biological discovery. The identification of RNA-repeat arrays associated with reverse transcriptases is genuinely intriguing and merits further investigation.

Read as written, that endorses the direction and stops short of endorsing a result. It is a fair thing to say about a structure whose function is unknown, and it is worth not rounding up.

One boundary Anthropic drew explicitly: Claude did not operate laboratory equipment. Amodei said that it may eventually be possible for Claude to run experiments itself by controlling lab equipment autonomously, with appropriate safeguards, but that the company is not doing that today.

What this changes for people building agents

Not prompting technique. What the run demonstrates is orchestration at a scale most teams do not attempt, with the parts that make long searches survivable made explicit: a bounded budget, a role split that keeps a multi-hour campaign from drifting, a tournament to rank what comes back, and a curator to deduplicate. The deliverable was not an answer. It was about twenty reports a human could actually read, distilled from a billion-scale search space.

That framing — agents as a funnel that compresses an unreadable space into a reviewable shortlist — travels well past biology, and it is a more useful takeaway than the headline. The cost anchor is useful too: 210 million tokens is a real but unremarkable spend at current frontier pricing. The scarce inputs were the design of the search and the human bench work that had to follow it. If ART's function turns out to be nothing much, the pipeline still stands. If it turns out to be programmable, the 21 hours will be the number everyone quotes, and the months of wet-lab validation still ahead will be the part they forget.