Anthropic published a result on 24 September 2026 that is genuinely new: Claude agents identified a previously undescribed enzyme system in bacteriophages, and human scientists then confirmed it exists in a laboratory.
It is also a result with a large hole in the middle, which Anthropic states rather than hides. Nobody knows what the thing does yet.
What Was Found
The system is called array-associated reverse transcriptases, or ART. It has three parts:
- A reverse transcriptase enzyme
- A partner gene
- A long array of evenly spaced DNA repeat sequences
That third component is why this got attention. Evenly spaced DNA repeats are the structural signature of CRISPR, the bacterial immune system that became the basis of modern gene editing. Finding the same architectural pattern in bacteriophages - the viruses that infect bacteria - is the part that made specialists sit up.
It is worth being precise about what that does and does not mean. A structural resemblance to CRISPR is not a claim that ART edits genes, or that it will. It means the shape is familiar and the shape is usually there for a reason.
The Search, By the Numbers
Anthropic published the scale of the run, which is the part most relevant to anyone evaluating agents for real work:
- ~950 agents running in parallel
- 21 hours of wall-clock time
- 210 million tokens consumed
- 200,000+ reverse transcriptases gathered
- 3,500 candidate systems identified
- 20 most compelling candidates analysed in depth
That funnel is the actual finding for most readers. 200,000 down to 3,500 down to 20 is a filtering job, not a reasoning breakthrough - and filtering at that scale is exactly what a swarm of agents is good at and what a research group with finite postdocs is not.
210 million tokens is also a number you can price. At current frontier rates that is a few hundred dollars of inference to sift a space no lab was going to sift by hand.
The Wet Lab Part
This is what separates the result from a press release. Human scientists expressed the protein in laboratory strains and ran biochemical and structural characterisation on it. Early experiments showed the ART array is also expressed as a set of distinct short RNAs.
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 of MIT and the Broad Institute - one of the people who turned CRISPR into a tool - called it "an exciting example of how AI agents can contribute to biological discovery." That is a careful sentence from someone who knows exactly how careful to be, and it is worth reading at the temperature it was written.
What Is Missing
Anthropic's own words: "we don't yet know its function," and work to understand it is ongoing.
That is the whole gap. A novel system has been found, confirmed to exist and confirmed to be transcribed. What it is for - immunity, mobility, something with no analogue yet - is unanswered. Until that lands, ART is a discovery rather than a tool.
Anthropic leading with the uncertainty rather than burying it is the right call, and unusual enough to note.
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 claimed Claude understood phage biology. It gathered 200,000 sequences, applied consistent criteria across all of them, and surfaced a shortlist small enough for people to work through. The interesting judgement - which 20, what to express, how to characterise it - stayed human.
That division of labour is available to any research group with a large search space and a small team, and it is the part of this story that generalises. The enzyme may or may not turn out to matter. The workflow already does.
What To Watch
- Function. Until ART's purpose is known, this is a well-documented curiosity. That answer decides whether it becomes a tool.
- Reproduction. Whether other groups can run a comparable agent sweep on a different sequence space and surface something real is the test of whether this is a method or a one-off.
- The cost curve. 210M tokens was expensive a year ago and is not now. Sweeps of this shape get cheaper every quarter, which changes who can afford to run them.
Sources