950 Claude Agents Scan 1.9 Billion Protein Clusters in 21 Hours and Find a Suspected Novel CRISPR-like Enzyme System—But All 10 Reruns Fail

On September 23, 2026, Anthropic announced a biological discovery: about 950 Claude agents, over 21 consecutive hours, autonomously scanned a DNA database containing about 1.9 billion protein clusters and identified a previously undescribed enzyme system in jumbo phages. The system was named ART (array-associated reverse transcriptases), and its DNA repeat array arrangement closely resembles that of CRISPR arrays. This was the first public instance of a frontier AI model completing the full chain from computational discovery to wet-lab confirmation.

According to Anthropic's official blog, the ART system consists of three components: a reverse transcriptase (RT), an adjacent companion gene, and a string of evenly spaced DNA repeat sequences. The underlying RT enzyme had been mentioned in prior research, but Claude was the first to notice the combination of all three occurring together—precisely this "holistic pattern recognition," rather than the discovery of a single element, formed the core of the scientific contribution claimed here.

What the AI Did and Did Not Do

The Claude agents were given a high-level direction—"explore bacterial DNA datasets and identify unknown protein families"—and ran autonomously without a specific hypothesis. According to Anthropic, the agents autonomously flagged patterns across massive numbers of protein clusters, generated candidates, narrowed the field, and ultimately delivered the ART system to human scientists for wet-lab validation.

This marks an important dividing line from AI's role in biology over the past years—protein structure prediction (AlphaFold), drug molecule screening. Those efforts were essentially still "given a structure/target, AI performs computation"; here, AI accomplished "hypothesis generation about patterns in an unknown space" and triggered follow-up validation in a physical laboratory. Stanley Qi, an associate professor of bioengineering at Stanford University, commented that AI identified "aberrant biological patterns that were previously difficult to detect and systematically tracked them as a research question."

The Most Critical Number: 10 Reruns, 0 Hits

However, one detail was glossed over lightly by some reports, yet it is key to understanding the true quality of this discovery: Anthropic reran the same search task 10 times, and each time it failed to find the ART system's repeat array pattern again.

What does this mean? It does not necessarily mean the discovery is invalid—cases in the history of science where a single discovery is difficult to systematically reproduce are not rare, especially in computational searches, where the randomness of sampling paths can make conclusions highly sensitive to initial conditions. But it does reveal an unresolved methodological problem: if an AI system's "discovery capability" can be reproduced only once under the same inputs, is it actually performing a robust scientific search, or did it make a lucky pattern match among 1.9 billion candidates?

For any institution planning to deploy AI agents at scale in life science research, this reproducibility problem is not a detail but a class of methodological challenge. Anthropic has not yet publicly explained how it will address this issue.

The Real Disagreement in the Scientific Community

Feng Zhang, a professor at MIT and the Broad Institute and a pioneer in CRISPR gene editing, said after reviewing the preprint that "the discovery of an association between reverse transcriptases and RNA repeat arrays is genuinely interesting and worthy of further study," and expressed hope that the work would "encourage more scientists to explore how AI can support their research."

This wording has a standard meaning in academia: it is not "we have discovered the next CRISPR," but "this is a lead worth exploring." Feng Zhang did not say the system can edit genes, nor did he say it has therapeutic potential.

Kevin Blake, a microbiologist at the University of Washington School of Medicine, was more direct. According to Al Jazeera, he noted that because the ART system shows "CRISPR-like" features, some people quickly interpreted it as "the next CRISPR," but there is no evidence that the system is a competitor to CRISPR or that it can be developed into a therapeutic application. He also noted that millions of bacterial species have not yet been studied and that countless CRISPR-like sequences remain to be cataloged, so the discovery of a new CRISPR-like structure is not itself a rare event.

These two reactions are not contradictory—together they depict the true coordinates of this discovery: a credible preliminary finding, but still a considerable distance from the narrative of a "new gene-editing tool."

Anthropic's Strategic Logic

Placed in a larger context, the timing of this release is quite telling. In spring 2026, Anthropic formed an internal biology research team in the San Francisco Bay Area, with an accompanying independent wet lab, and simultaneously pursued full-chain R&D from training Claude's biological capabilities to actual laboratory validation. This is the first public case of an AI company internalizing scientific discovery capability as part of its core product strategy, rather than licensing its model to external research institutions.

Dario Amodei has publicly claimed that AI could cure most diseases within 5 to 10 years. Regardless of how the scientific significance of the ART discovery is ultimately settled, it constitutes the first "deliverable evidence" for this grand narrative. More importantly, it sends a signal to the entire industry: AI agents can autonomously complete biological hypothesis generation in an unsupervised state and trigger validation in real physical laboratories—this pipeline is already running.

At the competitive level, this also exerts invisible pressure on OpenAI, DeepMind, and other organizations. AlphaFold has already reshaped the field of protein structure prediction; moving from "structure prediction" to a closed loop of "active hypothesis generation + experimental validation" is the next competitive battleground for AI in the life sciences. Anthropic's release this time may well be an entry declaration announcing that it has entered this race.

Independent Judgment

The discovery of the ART system is scientifically real: a documented novel molecular combination, confirmed by human wet-lab work and given preliminary affirmation by fellow researchers. But there are currently two substantive unknowns: First, the function is unknown—what ART actually does, whether it is programmable, and whether it has gene-editing potential cannot currently be answered; Second, reproducibility is in doubt—all 10 reruns failed, meaning the search method this discovery relies on has not yet reached a level of robustness that can be engineered.

This does not prevent ART from being a scientific lead. But the gap between the narratives on media and social platforms—"AI replaces traditional scientists" or "a new CRISPR has arrived"—and the cautious wording in Anthropic's own blog should draw the attention of rational readers.

Whether Anthropic can next systematize this discovery pipeline—making AI agents' biological hypothesis generation repeatable, batchable, and verifiable—determines whether it can change the paradigm of life science research.