Anthropic's Wet Lab Claims Its First Discovery: Claude Agents Found a CRISPR-Like Enzyme System ("ART")
Update (September 23, 2026): the life-sciences bet (Anthropic's Wet Lab Claims Its First Discovery: Claude Agents Found a CRISPR-Like Enzyme System ("ART")) has produced its first concrete artifact. Anthropic announced "Claude discovers a novel enzyme system with CRISPR-like repeats" — the debut result of its new life sciences research group and Bay Area lab. Claude agents, given only a high-level prompt, combed DNA databases and flagged a previously uncharacterized system in bacteriophages: a reverse transcriptase accompanied by an array of evenly spaced DNA repeats and an accessory protein — a layout reminiscent of CRISPR. Anthropic calls it ART (array-associated reverse transcriptases) and released a pre-print.
The scale of the autonomous run, from the post:
"After 21 hours spent searching this data by roughly 950 agents using 210 million tokens, one of the agents spotted something remarkable: a repeating pattern of DNA sequences that occurs next to the gene for an odd-looking RT."
And the moment of discovery, quoted verbatim from the agent's own analysis:
"[The DNA next to the RT] is spectacular: I can see by eye a tandem repeat array … that's a CRISPR-like … repeat array?!"
Independent validation is cautious, not conclusive. Feng Zhang — CRISPR pioneer, MIT/Broad — reviewed the pre-print: "This is an exciting example of how AI agents can contribute to biological discovery. The identification of RNA-repeat arrays associated with reverse transcriptases is genuinely intriguing and merits further investigation." Anthropic itself stresses the system's function is still unknown; the arrays are expressed as distinct short RNAs, "suggesting that something analogous may be at play" — nothing more. Coverage (Yahoo Finance/WION) echoed that "science isn't confirmed."
Capability-vs-hype audit, per the post's own disclosures: the agents gathered 200,000+ RTs, narrowed 3,500 candidate systems to 20, and produced human-readable reports — work Anthropic says would take an expert "weeks to months." But all lab work is performed by human scientists (BSL-1/BSL-2 only, no human-infecting pathogens), and Claude runs in the same Claude Science / Claude Code tools available to any scientist. The discovery is real as hypothesis-generation; the CRISPR comparison is a structural analogy, not a demonstrated function.
Why it matters: this is the first externally-checkable claim that an AI agent pipeline can originate (not merely assist) a novel biological-systems discovery — the concrete data point behind both the R&D-automation metrics (Anthropic's Automation Trajectory Goes to the IMF: Claude Projected to "Lead" 80–90% of Frontier Model Training by Next Year) and the biotech-acceleration framing in Amodei's pacing essay. It also sharpens the dual-use tension the same week's reporting highlighted: Anthropic's threat-intel report (Sept 10) detailed blocked bird-flu misuse attempts, while its own lab pushes AI-led biology forward. The function of ART, and whether the discovery holds up, is the testable next step.