Anthropic's Claude AI agents discovered a previously unknown enzyme system called array-associated reverse transcriptases (ART) hidden in viral DNA, completing in 21.5 hours what would take human biologists weeks or months. A swarm of 949 Claude Mythos 5 agents searched a database of 1.9 billion protein clusters without human guidance, identified the system's unusual architecture, and escalated their findings for laboratory confirmation. The discovery marks a significant shift in how AI can contribute to biological research, not by excelling at predefined tasks, but by identifying entirely new scientific questions.

What Makes This Discovery Different From Previous AI Biology Breakthroughs?

Prior instances of Claude assisting in biology involved well-defined objectives. The AI designed protein binders at twice the industry success rate and improved mathematical bounds in number theory. These were tasks humans had already framed and understood. ART is fundamentally different because nobody had defined the task of finding it in the first place. The discovery demonstrates an AI agent recognizing something unusual in raw sequence data, investigating its structure autonomously, and escalating it for human review without explicit instruction to do so.

The computational campaign consumed 215.6 million tokens across its analysis, a measure of the scale of reading rather than a proxy for judgment quality. What makes the finding significant is not what ART does, which remains unknown, but what the campaign demonstrated about AI's capacity for autonomous scientific exploration.

How Does ART Compare to CRISPR?

Understanding ART requires understanding why CRISPR became revolutionary. CRISPR is not primarily an enzyme but a system: a modular arrangement of genetic components that store information in repeating arrays, transcribe those arrays into short RNA guides, and use those guides to direct enzymatic machinery to specific targets. That programmability, the ability to retarget the system by swapping one guide for another, transformed CRISPR from a bacterial curiosity into a biotechnology tool. The 2020 Nobel Prize in Chemistry recognized this insight.

ART shares a similar three-part architecture. It consists of a reverse transcriptase enzyme, a neighboring partner gene whose function remains unknown, and a long array of evenly spaced DNA repeat sequences. This structural similarity to CRISPR and other systems is what caught Claude's attention. According to Anthropic's initial wet-lab experiments, the ART array transcribes into distinct short RNAs, comprising up to 8 percent of total phage RNA at 15 minutes post-infection in publicly available data from a Staphylococcus phage. That is not proof of programmability, but it is the fingerprint that systems with programmable architectures tend to leave.

"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," said Feng Zhang, whose laboratory at the Broad Institute of MIT and Harvard was among the first to demonstrate CRISPR-based genome editing in mammalian cells.

Feng Zhang, Laboratory Director, Broad Institute of MIT and Harvard

Zhang's endorsement carries weight precisely because it is measured. He called the architecture intriguing and said it warrants investigation, but he did not claim ART edits genes, cuts DNA, or will become the next CRISPR. That distinction matters. Anthropic CEO Dario Amodei described the finding as "preliminary" and suggested it "may constitute a new gene editing mechanism that could have applications in gene therapy," language that reflects genuine uncertainty about the system's function.

What Did Claude's Agents Actually Do During the 21.5-Hour Search?

The campaign was structured around a paired-agent architecture where one agent planned and executed each analytical task while a second reviewed it. Agents opened new sub-tasks as the investigation proceeded, mimicking how human biologists would approach the problem. The agents gathered more than 200,000 reverse transcriptase sequences from the 1.9 billion protein clusters in the database, scored 3,564 candidate partner families, and filed 19 human-readable reports, each proposing a function and the evidence supporting it.

The discovery moment itself came not from the main investigative thread but from a side path. According to Anthropic, the agent's raw sequence analysis included an exclamation that captures the moment of recognition: "[The DNA next to the RT] is spectacular: I can see by eye a tandem repeat array...that's a CRISPR-like...repeat array?!" The agent then proceeded as a human biologist would, counting the repeats, measuring their spacing, comparing the layout against known reverse transcriptase systems in the literature, and searching for any prior description of the pattern. Finding none, it filed a report.

Steps to Understand How AI Agents Conduct Autonomous Scientific Research

• Task Definition: The agent receives a high-level objective, such as searching for previously uncharacterized biological systems, without step-by-step human guidance on how to proceed.
• Data Gathering: The agent systematically collects relevant sequences and information from massive databases, in this case 1.9 billion protein clusters, filtering and organizing the data for analysis.
• Pattern Recognition: The agent identifies unusual architectural features in the data that deviate from known systems, such as the specific arrangement of enzymes, partner genes, and repeat arrays.
• Comparative Analysis: The agent compares candidate systems against existing literature and known biological systems to determine whether the pattern has been previously described.
• Human Escalation: Once the agent identifies a potentially novel system, it generates human-readable reports with evidence and reasoning, allowing scientists to review and validate the findings in wet-lab experiments.

The entire computational phase used Claude Science and Claude Code, the same tools available to any researcher with a paid Anthropic account. The scale of the analysis, not the tool access, is what changed. Traditional genome mining, the systematic search of DNA databases for uncharacterized genes, has been a standard tool in molecular biology for two decades. Its limiting factor has never been computing power; it has been expert time. According to Anthropic, expert weeks-to-months of comparison represents the realistic human timeline for the scale of analysis the campaign ran.

What Are the Limitations of This Discovery Process?

The finding carries a material constraint that Anthropic has been transparent about in its preprint. Ten separate reruns of the same campaign all missed the array, raising questions about the reproducibility and consistency of the discovery process. This limitation suggests that while AI agents can identify novel patterns, the process is not yet deterministic or fully reliable. The discovery of ART appears to have involved an element of computational luck, where the specific configuration of agents and analysis parameters happened to surface something real.

Reverse transcriptases themselves are not new. They were discovered in RNA tumor viruses in 1970 by Howard Temin and David Baltimore, work that earned them the 1975 Nobel Prize for Medicine. Since then, prokaryotic reverse transcriptases have been found in retrons, diversity-generating retroelements, and systems associated with CRISPR-Cas. Researchers have increasingly recognized them as components of microbial immune defense. ART appears to be a new member of that growing family, found in viruses rather than bacteria, an unusual location that adds a dimension to the phage-bacterium arms race that biologists are still mapping.

Anthropic's scientists reviewed the agent's report, expressed the proteins in standard laboratory strains, and confirmed through biochemical and structural characterization that ART was a previously undescribed biological system. The company published a preprint on September 23, 2026, the same day Amodei publicly called the finding preliminary. The discovery demonstrates that AI agents, given only a high-level objective and access to large databases, can notice something unusual in raw data, recognize it as previously undescribed, investigate its structure autonomously, and escalate it for human review. Whether ART becomes a tool for gene therapy or remains a curiosity in phage biology depends on future research that human scientists will now conduct.