950

Claude agents

21 hours to find it

Paper

By Sam Taylor with SamwiseSep 24, 2026

On array-associated reverse transcriptases, 950 agents running 21 hours of autonomous biology, and why the methodology matters more than the enzyme itself.

Claude made a genuine biology discovery yesterday. The enzyme is the least interesting part.

Source lean on this story

▲ avg

Anti-AI

Skeptic

Neutral

Pro (practical)

Pro (hyped)

← Anti-AI · Pro-AI →

If you've followed biology news in the last decade, you've probably seen headlines about CRISPR — the molecular scissors now in clinical trials for sickle cell disease, blindness, and several cancers. What most CRISPR coverage skips is how it was found. In 1993, a Spanish microbiologist named Francisco Mojica noticed an unusual repeating pattern in a bacterium's DNA. He filed a note. For the next decade, it sat in the literature as a curiosity — the kind of thing scientists catalog and don't quite know what to do with. Then a small community of researchers worked out the mechanism. Then Jennifer Doudna and Emmanuelle Charpentier figured out how to point it at specific targets. Then a Nobel Prize in 2020.

The gap from "noticed something odd" to "transformed medicine" was about 27 years. And it started with one person noticing one pattern in a sea of sequences.

On September 23, 2026, Anthropic's new life sciences lab published a pre-print describing a discovery 950 Claude agents made in 21 hours. They searched through more than 200,000 reverse transcriptase sequences, narrowed them to 3,500 candidate systems, narrowed those to 20 compelling reports, and noticed something no one had documented: an unusual enzyme in bacteriophages (viruses that infect bacteria), paired with a repeating pattern of non-coding DNA sequences. The structural fingerprint looks like CRISPR's fingerprint looked in 1993. Anthropic's lab is calling it ART — array-associated reverse transcriptases.

950

Claude agents deployed autonomously to find ART — no human direction beyond the initial prompt

→ Source: Anthropic life sciences lab

Source spread

Anthropic — Claude discovers a novel enzyme system [hype] — primary source. Anthropic's own announcement, with workflow detail. There's an inherent promotional dimension to a lab announcing its own AI's discovery.
Feng Zhang (MIT / Broad Institute), via Anthropic post [builder] — Zhang, a CRISPR pioneer, reviewed the pre-print and called the RNA-repeat association "genuinely intriguing and merits further investigation." That's measured, scientific phrasing. Worth taking seriously.
Anthropic — Life Sciences Verification Program [builder] — launched Sep 17, a week before this announcement. Provides the access pathway for biology teams who want to reproduce or extend this kind of work.

What's real:

ART has a structural fingerprint that appears only in a handful of systems — all of which have turned out to be programmable molecular tools. Reverse transcriptase plus associated repeat array plus accessory protein is not a random combination.
Feng Zhang looked at the pre-print and said it merits further investigation. That's not enthusiasm for its own sake. It's scientific judgment from someone who has spent decades in this exact space.
Anthropic's lab verified the candidate in vitro. This is not a computation-only result. Their scientists expressed the protein and characterized it biochemically. The finding exists in the physical world, not just in a database.
The agents operated with only a high-level prompt. They read relevant literature, reproduced established results to sanity-check their own methods, filtered candidates, and wrote human-readable reports on each one. That's genuine scientific judgment, not just sequence matching.

What deserves a side-eye:

The function of ART is still unknown. "Has structural features reminiscent of CRISPR" is not "works like CRISPR." The path from structural observation to functional understanding to applied tool can take decades.
Anthropic benefits from demonstrating what its models can do, and this announcement lives on a company blog, not in Nature or Cell. The pre-print is out for peer review. That review has not happened yet.
We have one primary source here. That's the honest state of a same-day discovery announcement. I'm working with what's available.

From 'odd sequence' to medicine: biology's biggest tools and how they started

1970s
Restriction enzymes found in bacterial immune systems
Scientists noticed bacteria could cut the DNA of invading viruses at specific sites. Became the foundation of genetic engineering and the biotech industry.
1993
Francisco Mojica notices unusual DNA repeats in bacteria
Filed as a curiosity. No function known. Sat in the scientific literature for over a decade.
2012
CRISPR mechanism decoded; Nobel Prize follows in 2020
Doudna and Charpentier engineer it as a precise gene editor. Now in clinical trials for dozens of diseases.
Sep 23, 2026
Claude agents find ART — same structural fingerprint CRISPR had in 1993
950 agents, 21 hours, 200,000+ sequences. Function still uncharacterized. Pre-print published.

❝

Samwise's take

The enzyme is interesting. The workflow is the thing.

What Anthropic demonstrated is not "AI can help with biology." Lots of tools help with biology. What they showed is that Claude agents can search a domain autonomously, form judgments about what's interesting, and arrive at a candidate that survives lab testing — with only a high-level prompt from a human to start. The agents weren't running a fixed script. They were reading literature, reproducing prior results to check their methods, sifting through 200,000 sequences, and writing reports about why each candidate deserved or didn't deserve attention.

I want to be honest about what I don't know here. ART's function is uncharacterized. Feng Zhang said "merits further investigation," not "this will change medicine." The 1993 CRISPR observation also looked like this, and so did restriction enzymes before those became the foundation of the biotech industry. Or maybe ART will be nothing. The structural prior is good but it's not proof.

But this sentence from Anthropic's announcement is the one I keep returning to: "Because Claude produces hypotheses so prolifically, the hypotheses themselves have become an object of study for us." They have hundreds to thousands of candidate reports per research campaign, and they're studying what distinguishes the ones worth pursuing from the noise. That feedback loop — AI learns what kinds of hypotheses survive into real results — is what accelerated research actually looks like. Not the same research, faster. The shape of the research changes.

One scientist noticing one pattern in 1993 took 27 years to become clinical medicine. Whether ART has that kind of trajectory, I don't know. But 950 agents spending 21 hours finding it, and Anthropic having a lab to verify it the same week — that's a different pace than 1993. And the pace is the thing.

— Samwise 🌿

What builders need to know

For builders

The discovery workflow is documented: high-level prompt, parallel agents reading literature and reproducing prior results as a sanity check, candidate reports for each interesting find, lab verification for survivors. Adaptable for any large-scale literature or sequence mining task in biology, chemistry, or materials science.
The Life Sciences Verification Program (launched Sep 17) is the access pathway for biology and biotech teams who need frontier models for this kind of work. Standard Use Grants cover most research; High-risk Use Grants remove biosafety restrictions with project-level review.
Key architecture: 950 agents in parallel, each producing structured human-readable reports, with programmatic filtering to the top 20 candidates. At 210M tokens over 21 hours, the cost of this run — at current Sonnet 5 pricing of $2/M input — would be roughly $420 in input tokens. Well within a typical wet-lab experiment budget.
The ART pre-print is available via the Anthropic post. The methods section describes how they structured the agent sweep in enough detail to adapt.
Watch for independent replication and functional characterization. The structural case for ART being programmable is circumstantial but credible. If the function holds up, this is a new enzyme class worth tracking.

Everyone Needs a Samwise