When a discovery starts with a scientist noticing something odd in a DNA database, the bottleneck is rarely the noticing itself. It’s the weeks of manual analysis that follow: reading literature, reproducing results, filtering candidates, and writing reports. Anthropic’s new life sciences lab just showed what happens when you hand that bottleneck to a fleet of agents.
In a preprint released September 23, 2026, Anthropic describes how Claude discovered a novel enzyme system with CRISPR-like repeats after 21 hours of searching by roughly 950 agents using 210 million tokens. The system, called array-associated reverse transcriptases (ART), consists of a reverse transcriptase, a partner gene, and a long array of evenly spaced DNA repeats. The repeat layout resembles a CRISPR array, and early experiments show the ART array is expressed as distinct short RNAs — a pattern that has only been found together in a handful of programmable systems.
The workflow that found it
Anthropic’s scientists gave Claude a high-level prompt: search a massive DNA sequence database for interesting reverse transcriptases. The agents then gathered over 200,000 RTs, picked out 3,500 new candidate systems, and narrowed those to 20 compelling candidates with human-readable reports. For an expert scientist, this analysis can take weeks to months.
One agent noticed an unusual RT family and examined it further. While reading the raw DNA sequence, it flagged a tandem repeat array next to the RT gene. It counted repeats, measured spacing, compared the layout to known systems, and searched literature for prior reports. Then it filed a report for human review. The lab confirmed the pattern experimentally.
This is not a one-off. Anthropic describes a repeatable pattern: Claude reads relevant literature, reproduces established results from public data, searches for genomic neighbors that fit no described system, and writes short reports proposing functions with supporting evidence. Most candidates are eliminated in follow-up analysis. A survey may end with one candidate worth testing, or none.
Why the hypotheses themselves matter
Because Claude produces hypotheses so prolifically, Anthropic’s team now studies the hypotheses as an object. With hundreds to thousands of candidate reports from a single campaign, they ask what distinguishes proposals worth testing from those set aside. What they learn goes back into the instructions given to Claude, teaching it to mimic scientific taste.
This feedback loop is the real product insight. The discovery of ART is interesting, but the durable asset is the workflow that turns raw sequence data into a ranked list of testable candidates. That workflow is built on the same tools any scientist can use: Claude Science and Claude Code, plus a custom harness that coordinates many Claude sessions running in parallel.
For builders thinking about agent-driven research, the pattern echoes what we’ve seen in other domains: the value is not in a single model call but in the loop between generation, evaluation, and refinement. Anthropic’s approach to hypothesis triage is similar to how third-party AI assessments are changing what builders prioritize — the hard part is deciding what to test, not generating options.
What’s still open
Anthropic does not yet know the primary function of ARTs. The system is found mainly in bacteriophages, and experiments are underway to determine how it works. The lab operates at BSL-1 and BSL-2 only, and all lab work is performed by human scientists. The preprint is available for the community to review.
Feng Zhang, a CRISPR pioneer at MIT and the Broad Institute, called the identification of RNA-repeat arrays associated with reverse transcriptases “genuinely intriguing” and said it merits further investigation.
For product builders, the takeaway is not that Claude found a new enzyme. It’s that a well-scoped agent workflow can compress weeks of expert analysis into hours, and that the workflow itself becomes a reusable asset. The next step is watching whether other labs adopt this pattern — and whether the hypotheses Claude generates hold up under experimental scrutiny.
Sources
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.