On September 23, 2026, Anthropic announced that agentic Claude searched 200,000+ reverse transcriptases in a DNA database and surfaced a novel enzyme system, later verified in the lab.
- This guide provides comprehensive, actionable information
- Consider your specific workflow needs when evaluating options
- Explore our curated AI Assistants tools for specific recommendations
What actually happened
On September 23, 2026, Anthropic announced that its newly formed life sciences research group had used agentic Claude to identify a novel biological system in bacteriophages: array-associated reverse transcriptases (ART), made up of a reverse transcriptase gene, a partner gene, and repeating DNA sequences that resemble CRISPR arrays. The prompt Claude was given was broad and largely hands-off: search a massive DNA sequence database for interesting new examples of reverse transcriptases. Everything past that instruction — the search strategy, the narrowing, the candidate selection — ran without step-by-step human direction.
That is the real, verifiable part of the story. What ART actually does is not yet known, and Anthropic says so directly.
What was actually verified, and by whom
Claude's role was search and candidate selection, not confirmation. After the agents narrowed 200,000+ sequences to 20 compelling candidates, Anthropic's own lab scientists expressed the proteins in standard laboratory strains and ran the biochemical and structural characterization by hand. Feng Zhang of MIT and the Broad Institute, a CRISPR pioneer, reviewed and commented on the findings independently of Anthropic. That combination — AI-driven search at a scale no human team searches at, followed by conventional wet-lab verification — is the actual model here, not "AI discovers biology alone."
The honest caveat Anthropic itself makes
Anthropic's own writeup is unusually direct about what this is not: "Although we don't yet know its function," and "our work to understand the primary function of ARTs is ongoing." The company says it is sharing early findings specifically to demonstrate agentic search capability, not to announce a confirmed application. ART was found in bacteriophages, not in a context with an obvious immediate medical use, and Anthropic notes that AI acceleration was "less conducive" to the molecular biology lab work itself compared to other parts of the process — the humans, not the model, did the slow part. The plausible upside, per the researchers, is that ART may turn out to be programmable in a CRISPR-like way (cutting, copying and pasting DNA), based on the fact that its array is also expressed as distinct short RNAs. That is a hypothesis to test, not a result.
Why the scale is still the real news
Searching 200,000 sequences for candidates that look interesting is exactly the kind of task that does not scale for a human research team: it is high-volume, pattern-matching-heavy, and mostly a filter for the small number of things worth a scientist's time. 950 agents running 21 hours to get from 200,000 down to 20 is a genuine capability difference from a lab doing the same triage by hand over weeks or months. That's the part of this story that generalizes past biology — the same agentic-search-then-human-verification pattern shows up in coding, research synthesis and other high-volume triage tasks; see agent swarms and research agents for how that scaling pattern plays out elsewhere.
Who should actually care
For how agent orchestration at scale works in other domains, see agent swarms and research agents. For the model family behind this, see Claude and every AI assistant in the directory.
Explore curated tools related to this guide: