Claude Discovers a Novel Enzyme System: What 950 Agents Actually Did

About 950 Claude agents spent 21 hours and 210 million tokens reading public DNA data, and one of them flagged something nobody had written up before. Anthropic says Claude discovers a novel enzyme system this way: a reverse transcriptase sitting next to a long array of evenly spaced DNA repeats, a layout that resembles CRISPR. Anthropic announced the result on 23 September 2026 alongside a new life sciences research group and lab, and released a pre-print.

The headline is easy to overstate, so it helps to be precise about what happened, what humans did, and what is still unknown.

What Claude found

Reverse transcriptases (RTs) are enzymes that copy RNA into DNA. In bacteria, many of them act as part of the immune system, and most known RT families were found through genome mining: searching sequence databases for uncharacterized genes and working out which ones are unusual.

According to Anthropic, the agents gathered over 200,000 RTs, picked out 3,500 new candidate systems, and narrowed those to the 20 most compelling, each written up as a human-readable report. One RT family, found mainly in bacteriophages (viruses that infect bacteria), had three parts: the RT, a partner gene beside it, and a long array of repeat sequences. Anthropic's team named it array-associated reverse transcriptases, or ART.

Two details matter here. First, the RT itself was not new. Anthropic says it had been identified in previous studies, found in a jumbo phage. What Claude appears to be the first to notice is the surrounding system: the repeat array and the accessory protein. Second, the lab work so far shows that the array is expressed as a set of distinct short RNAs. That is suggestive, because CRISPR arrays also hold a bank of short RNAs, but it is not a function.

Anthropic says so directly: "we don't yet know its function." Further experiments are underway.

How the agents worked, and where humans stepped in

The division of labour is the most useful part of the announcement for anyone building agent systems. Anthropic describes its involvement as "limited to the initial prompt and the lab work." Everything between those two points was agents.

The workflow it describes follows a pattern worth copying:

1. Reproduce known results first. Claude reads the literature on a protein family and reproduces established results from public data, to check its own methods before looking for anything new.
2. Search for misfits. It then looks for family members or genomic neighbours that fit no described system.
3. Write a report per candidate. Each candidate gets a short report proposing a function and laying out the evidence.
4. Critique its own output. In follow-up analyses Claude evaluates that evidence, and Anthropic says most candidates are eliminated at this stage. A survey can end with one candidate worth testing, or none.
5. Humans test survivors. Scientists express the protein in lab strains and characterize it, with Claude helping interpret the data. All bench work is done by people.

The team says it runs this in Claude Science and Claude Code, "the same tools available to any scientist," sometimes with its own harness that coordinates many parallel sessions. That is the same shape as the parallel-thread setup described in Claude Code Projects, scaled to hundreds of agents.

AI agents and scientific discovery: what is actually new

There have been plenty of claims about AI agents in scientific discovery this year, many of them about solving problems people had already posed. This one is different in kind. Nobody asked Claude to explain a known system. It was pointed at a large, public dataset and came back with a candidate nobody had flagged.

The mechanism is not cleverness so much as coverage. Anthropic notes this type of analysis can take an expert "weeks to months." The agents did the reading in 21 hours, at a scale no small team would attempt, and then filtered hard. The same logic sits behind Google's multi-agent bug pipeline covered in Google's AI fixed 1,072 Chrome security bugs: the value comes from agents working through a backlog humans never had time to open.

Feng Zhang, one of the pioneers of CRISPR genome editing and a professor at MIT and the Broad Institute, reviewed the pre-print and called it "an exciting example" of agents contributing to biology, while describing the finding as something that "merits further investigation." That is a fair summary of where it stands.

The part Anthropic is studying: too many hypotheses

One line in the announcement deserves more attention than the enzyme. Because Claude "produces hypotheses so prolifically," the team says the hypotheses themselves have become an object of study. With hundreds to thousands of candidate reports per campaign, they are asking what separates proposals worth testing from ones they set aside, and feeding that back into Claude's instructions so it can "mimic our own scientific taste."

This is the real bottleneck in most multi-agent research workflows, scientific or not. Generating candidates is cheap. Deciding which ones deserve a human's time is not. Anthropic's answer is to treat human judgement as training signal for the filter, rather than as a final manual review of everything.

What this means if you build with agents

Most readers are not mining genomes, but the pattern transfers to any large pile of records your team has never had time to read properly: support tickets, contracts, call transcripts, old research.

• Make agents prove their method on known cases first. Reproducing established results before searching for new ones is a cheap check against confident nonsense.
• Require a written report per candidate. A short, evidence-backed write-up is something a human can review quickly. A score is not.
• Build the elimination step in. The workflow is designed to kill most candidates. If your pipeline surfaces everything it finds, it will bury the reviewer.
• Keep humans on the expensive, irreversible step. Here that is lab work. In a business it might be sending, signing, or spending.
• Be honest about what is unconfirmed. Anthropic published before knowing what ART does, and said so. That is the right way to report an early result.

What to watch next

The open question is function. "Claude discovers novel enzyme" is accurate as far as it goes, but the enzyme's job is still a blank. If ART turns out to be programmable in the way CRISPR is, this becomes a significant biology story. If it does not, it is still a working demonstration of agents doing the first, most tedious stage of discovery on their own. Anthropic says it wants to work with outside scientists on other research questions, so more results from this lab are likely.

Sources: Anthropic, "Claude discovers a novel enzyme system with CRISPR-like repeats" (23 Sep 2026) and the linked technical pre-print.