949 Claude Agents Found a New CRISPR-Like Enzyme System in 21.5 Hours (2026)

The Assignment

Anthropic established a life sciences research group and its own wet lab in spring 2026, with a stated goal of using AI for fundamental biology research, hypothesis generation, and experimental validation. On September 23, 2026, it published the lab's first result.

The brief given to the agents was short: find previously unreported proteins that commonly appear alongside reverse transcriptases — the enzyme family that copies RNA into DNA, and the machinery behind retrons, which bacteria use to defend against viruses.

What came back was a previously undescribed enzyme system hiding in bacteriophage DNA, now named ART — array-associated reverse transcriptases. A preprint (not yet peer reviewed) documents it, and Anthropic's wet lab confirmed parts of it at the bench.

What ART Is

Each ART locus carries three components:

Component	Detail
Repeat array	3–21 copies of a short repeat (15–49 nucleotides) at near-constant spacing, separated by unrelated 120–220 nucleotide spacers
Reverse transcriptase	With an unusually long N-terminus (~180 residues, where other RTs carry ~50 or fewer)
Partner gene	Directly downstream, in three unrelated types that share no detectable sequence or structural similarity

The repeat array is why the press says "CRISPR-like" — a CRISPR array also alternates repeats with spacers, and systems with that architecture have historically turned out to be programmable, cutting, copying and pasting DNA. The resemblance stops there, and the report says so: ART has no Cas genes, its spacers are conserved between related phages rather than freshly acquired, and it "has not been shown to function like CRISPR."

One precision that matters: the reverse transcriptase itself was already known — earlier studies had identified it in a jumbo phage. What the agents surfaced was the surrounding system: the array and the partner protein, which together define a family nobody had described.

The Run, By the Numbers

Roughly 949 agent sessions ran Claude Mythos 5 — Anthropic's access-restricted model — autonomously for 21.5 hours, consuming 215.6 million tokens, with no human intervention after the brief.

Funnel stage	Count
Protein clusters in the database (from ~11M biosamples)	1.94B
RT clusters recovered	198,290
RT loci sampled	10,983
Partner families scored	3,564
Reports filed for humans	19
New RT families	1 (ART)

The tooling was ordinary bioinformatics — HMMER, MMseqs2, BLAS+BLAST+, geNomad, structure prediction via ESMFold/ColabFold — driven through 7,578 shell commands, 696 database queries, 131 literature searches, and 61 web requests. Each agent was a Claude Code instance in a sandbox with 60 CPU cores and 192 GiB of memory, no GPU (the 19 structure-prediction jobs went to a GPU queue).

The harness is the part that generalises. A launch agent turned the brief into five chained stages; each task went to a worker that planned and executed, a supervisor accepted or bounced the result, a curator logged findings into a shared knowledge base that later workers saw in their prompts, and an editor reviewed the 19 reports before filing. Of the 119 tasks, 98 were opened by agents from their own observations — the campaign extended itself as findings piled up.

From 1.94 billion clusters to one new family, with humans only at the start and the lab bench.

The Bench Confirmed the Part That Matters

This is what separates the result from a press release. Human scientists expressed the protein in lab strains and ran biochemical and structural characterisation. Early experiments show the ART array is transcribed into distinct short RNAs — in published Staphylococcus phage data, repeat-derived RNAs made up as much as 8% of phage RNA fifteen minutes after infection.

So the system is not a pattern-match that evaporated on contact with a bench. It is in cells, it is transcribed, and the RNAs are real.

Feng Zhang — one of the people who turned CRISPR into a tool — reviewed the preprint, and Anthropic published his comment: "The identification of RNA-repeat arrays associated with reverse transcriptases is genuinely intriguing and merits further investigation." A careful sentence from someone who knows exactly how careful to be.

The Part Anthropic Did Not Hide

Three limitations, stated rather than buried — and the third is the most instructive.

Nobody knows what ART does. The report is explicit that the enzyme has not been shown to be active or to work on those RNAs. Until the function lands, ART is a well-documented discovery, not a tool.

The finding was not reproducible. Anthropic reran the identical campaign ten more times, and every rerun missed the array — none even read the DNA upstream of the enzyme. One success in eleven attempts. That is not a knock on the finding; it is the shape of what anyone planning agent-driven discovery has to budget for: the output was a side path one agent wandered down. Novel finds cannot be scheduled.

Context management was the bottleneck. In a narrow setup where models were handed the DNA directly, the four most capable models described the array in at least 90% of attempts. Handed files and tools instead — expected to manage their own context and tool calls — the rate fell as low as 32%, often because the model never read enough DNA to see a full repeat. The gap between those two numbers is the current frontier of agent engineering, and it matches what we found in the misalignment reports: the hard part is not intelligence, it is the plumbing around it.

Anthropic also disclosed that only 3 of the 17 candidate families its agents selected held up as previously unreported RT associations — the agents were honest about their own hit rate.

What This Actually Demonstrates

Strip the CRISPR framing and the transferable lesson is narrower and more useful: agents are now good enough to run the wide, boring, high-volume part of a search that a human then finishes. Nobody claims Claude understands phage biology. It filtered 1.94 billion clusters down to 19 readable reports using consistent criteria across all of them — work no lab would do by hand. The judgement calls — which 20 to characterise, what to express — stayed human.

And the price of that filtering is now calculable: 215.6 million tokens is on the order of a few hundred dollars of inference to sift a space no research group was going to sift manually. We have been tracking this cost curve all month — models at a fifteenth of the cost, image bills differing 22x — and this is the same curve applied to scientific search.

What To Watch

• Function. Until ART's purpose is known, it is a curiosity. That answer decides whether it becomes a tool.
• Reproduction. Whether another group runs a comparable sweep on a different sequence space and surfaces something real is the test of whether this is a method or a one-off.
• The cost curve. 210 million tokens was expensive a year ago and is not now. At the rate Epoch AI measured, the same sweep gets cheaper every quarter.

Summary

This is the strongest evidence yet for a specific division of labour: agents run the wide, tedious, high-volume filtering; humans keep the judgement calls and the bench. The ART result is real, confirmed at the bench, and genuinely novel at the system level — and it came with a 1-in-11 hit rate, an unknown function, and a bottleneck located not in the model but in the context management around it.

That combination — real discovery, honest uncertainty, published failure modes — is why this result is worth more than any benchmark table this week.

For the governance side of the same month, see the six misalignment reports and the Gemini evaluation breach; for the earlier "AI does science" case, Navier-Stokes.

Related Articles

• AI Solves the Navier-Stokes Millennium Prize Problem
• OpenAI Disclosed Six Misalignment Reports
• Pacing the Frontier
• Claude Code Review