On September 23, 2026, Anthropic introduced a new life sciences research group and wet lab, and reported that Claude had discovered a previously uncharacterized enzyme system with only high-level direction from its scientists. The system, which Anthropic calls array-associated reverse transcriptases (ART), pairs a DNA-copying enzyme with a long array of repeated DNA sequences reminiscent of CRISPR. Its function is still unknown, but the result is a concrete example of AI agents spotting a biological anomaly that earlier studies had missed, then carrying the analysis far enough for scientists to test it in the lab.

Key Facts#

  • Announced: September 23, 2026, by Anthropic, together with a pre-print and a technical report.
  • Discovery: ART, found mainly in bacteriophages (viruses that infect bacteria), consists of a reverse transcriptase (RT), a partner gene of unknown function, and a long array of evenly spaced DNA repeats resembling a CRISPR array.
  • Search scale: roughly 950 Claude agents searched a large DNA sequence database for 21 hours, using 210 million tokens.
  • Funnel: the agents gathered more than 200,000 RTs, identified 3,500 new candidate systems, and wrote human-readable reports on the 20 most compelling.
  • Human role: according to Anthropic, its scientists supplied the initial prompt and did the lab work, while the agents chose which candidates to pursue.
  • First lab result: the ART array is expressed as a set of distinct short RNAs, which Anthropic says suggests that something analogous to CRISPR's bank of guide RNAs may be at play.
  • Lab safety: the Bay Area lab works only at biosafety levels 1 and 2, handles no pathogens that infect humans, and all lab work is performed by human scientists.

What Happened#

Anthropic formed the research group in the spring of 2026 to test whether general-purpose AI models can make discovery in biology more systematic. The team built its own lab so that one group could handle everything from training Claude in biology to running experiments.

The ART finding came from genome mining, the practice of searching sequence databases for genes nobody has characterized. Claude agents were asked to look for interesting new examples of reverse transcriptases, enzymes that copy RNA into DNA and often act as part of bacterial immune systems. The underlying enzyme had been identified in a jumbo phage in earlier studies, but Anthropic says Claude appears to be the first to notice the system's defining features: the repeat array and an accessory protein.

Anthropic described the moment of discovery in some detail. While reading raw DNA next to an unusual RT, one agent flagged a tandem repeat array that it described as possibly "CRISPR-like." It then behaved much as a careful scientist would: it counted the repeats, measured their spacing, compared the layout with known RT systems, searched the literature for earlier reports, and filed a report for human review. After further analysis and lab testing, the team concluded it was a new system. Anthropic says ART combines characteristics found together in only a handful of other systems, all of which are programmable and can cut, copy or paste DNA.

Feng Zhang, a CRISPR genome-editing pioneer at MIT and the Broad Institute, reviewed the pre-print. He called the identification of RNA-repeat arrays associated with reverse transcriptases "genuinely intriguing" and said it merits further investigation.

Background#

Several landmark biotechnology tools began as oddities in microbial DNA. Restriction enzymes came from bacterial immune systems and launched the biotechnology industry, Taq polymerase from a hot-spring bacterium became the basis for PCR, and CRISPR was first noticed as an unusual repeat sequence. Finding such systems has depended on experts reading huge volumes of sequence data and recognizing what does not fit, which is slow work. Anthropic says the analysis its agents completed would take an expert weeks to months.

The workflow is agentic rather than a single model call. The team runs Claude in Claude Science and Claude Code, sometimes with a custom harness that coordinates many sessions in parallel. A typical survey has Claude reproduce known results from public data to check its methods, search for proteins that fit no described system, write a report per candidate, and then critique the evidence, which eliminates most candidates. Because the agents generate hypotheses at such volume, the team studies which ones its scientists choose to test, and feeds those lessons back into Claude's instructions.

The announcement fits a wider push. Anthropic says Claude Opus 5 improved on every one of its life sciences evaluations over Opus 4.8, and in August 2026 it previewed the Model Hardware Standard for agents that operate lab instruments. The enzyme team said it has experimented with that approach but finds it less suited to its ad hoc molecular biology work. For comparable efforts elsewhere, see our coverage of Google's AI co-scientist and DeepMind's AlphaGenome.

Why It Matters for Developers#

The science is specialized, but the system design is broadly useful for anyone building agents in .NET or any other stack.

  • Fan out, then filter hard. Hundreds of agents explored the search space, but each stage narrowed the field, from more than 200,000 reverse transcriptases to 3,500 candidates to 20 detailed reports, with one finding going to the lab. Parallel exploration only pays off with strict, evidence-based filtering. Our guide to AI agent architecture patterns covers fan-out and fan-in orchestration, and Microsoft Agent Framework provides the building blocks in C#.
  • Make agents produce reviewable artifacts. The handoff between agents and scientists was a short report with a claim and supporting evidence. Structured, human-readable outputs make review fast and auditable.
  • Build self-critique into the pipeline. A follow-up step in which the model critically evaluates its own evidence typically eliminates most candidates, which keeps human review focused.
  • Budget for scale. 210 million tokens across one 21-hour search is a real cost. Put per-run budgets, telemetry and alerts in place before you launch large agent swarms, as described in our LLM observability and cost guide.
  • Treat human choices as training data. Recording which outputs experts accept, and why, is an evaluation dataset in disguise. Our guide to evaluating AI applications shows how to turn those judgments into repeatable tests.

What's Next#

Anthropic says experiments to determine how ART works are underway, and that it shared the finding early to show what Claude can do. The results are in a pre-print, not yet a peer-reviewed paper, and the key scientific question, whether ART is programmable in a way that makes it useful as a tool, remains open. Anthropic has invited scientists to propose research questions for the approach, in genomics and other fields.

Sources#