On September 23, Anthropic announced a new life sciences research group and an early result from it: Claude autonomously identified a previously uncharacterized enzyme system in bacteriophage genomes. The system, which Anthropic calls array-associated reverse transcriptases (ART), consists of a reverse transcriptase, a partner gene, and a long array of evenly-spaced non-coding DNA repeats. That repeat structure resembles CRISPR arrays closely enough that wet-lab experiments confirmed the arrays are expressed as distinct short RNA molecules — behavior that in CRISPR systems encodes the programmable recognition sequences that make gene editing work.
The discovery matters, but the method is arguably more interesting.
The compute budget was substantial but not outrageous: approximately 950 Claude agents ran in parallel for 21 hours, consuming 210 million tokens while working through roughly 200,000 reverse transcriptases, narrowing to 3,500 candidate systems, then to 20 worth proposing to human scientists. A grad student would have taken months to get to the same funnel. The agents didn’t need months.
What the pipeline actually looked like: agents surveyed existing literature on reverse transcriptase protein families, reproduced established classification results from public data to calibrate their readings, then identified genomic neighbors around each RT — the uncharacterized flanking genes that might indicate a functional system rather than an isolated enzyme. For each candidate, Claude generated a structured hypothesis report with supporting evidence. Human scientists then ran conventional wet-lab experiments on the flagged candidates to confirm whether the predictions held.
This is a sensible division of labor. The agents’ comparative advantage is scale: they can hold a lot of literature context and pattern-match across many protein families simultaneously. The human lab’s comparative advantage is verification: actually expressing proteins, running binding experiments, checking whether the RNA molecules the AI predicted are real. Neither part works without the other. Anthropic’s Bay Area lab operates only at BSL-1 and BSL-2 biosafety levels, meaning no human pathogens — a deliberate constraint that keeps the discovery work in a verifiable, relatively low-risk range.
Whether ART turns out to be biotechnologically significant is an open question. CRISPR’s programmability comes from the combination of a nuclease (Cas9 or similar) with guide RNAs encoded in the repeat array — and the ART system has arrays that express into short RNAs. Whether there’s an associated nuclease or other effector, and what it does, is what Anthropic says its ongoing experiments are trying to determine. They’ve also invited external research proposals, which is the appropriate posture when you’ve found a signal but don’t yet know what it means.
The claim being made here is measured: a novel, uncharacterized enzyme system was found by AI agents working through sequence databases, and confirmed by wet-lab experiments. That claim is real. The more ambitious framing — transformative biotechnology breakthrough — should wait until the function is understood. Right now Anthropic has found a pattern with CRISPR-like structural features. That’s a meaningful first step, not a conclusion.
The more structural implication is about how biology research changes when 950 agents can fan out across a public database for 21 hours at a cost that would have been a small lab’s annual compute budget five years ago. Sequence databases have been accumulating uncharacterized systems for decades, waiting for enough patient attention to work through them. That attention constraint is going away.