Anthropic has reported the first result from its new biology lab. A large group of Claude agents working largely on their own identified an uncharacterised enzyme system in bacteriophage DNA whose layout resembles CRISPR. The company announced the finding on 23 September 2026 in a blog post and a preprint that has not yet been peer reviewed. It calls the system array-associated reverse transcriptases, or ART.
What the system actually does is still unknown. The more concrete news is the method: an agent-driven search that went from databases to candidates for wet-lab testing in less than a day.
What Claude Found
Bacteriophages are viruses that infect bacteria. The ART system Anthropic found in them has three parts:
- A reverse transcriptase, an enzyme that copies RNA into DNA
- A partner gene located next to it
- A long array of evenly spaced DNA repeats, between 3 and 21 of them, laid out much like a CRISPR array
In CRISPR, the array stores a library of RNA guide sequences, and that is why CRISPR-Cas systems can be programmed as gene-editing tools. Anthropic's first experiments found that the ART array is also expressed as a set of distinct short RNAs, which suggests a similar mechanism could be involved. The company says it has not yet determined the system's main function.
How 950 Agents Did the Search
The scale of the computational work is the main technical result. By Anthropic's account:
- About 950 Claude agents ran for 21 hours
- They used around 210 million tokens
- They collected more than 200,000 reverse transcriptases from public sequence databases
- They flagged about 3,500 candidate systems and selected 20 for detailed reports
The work was organised as a pipeline. One agent planned and carried out each task, a second agent reviewed the output, and agents created new sub-tasks when they found promising leads. Anthropic says the important observation came when an agent studying an unusual reverse-transcriptase family noticed the repeat array directly in the raw DNA sequence, a pattern human researchers had missed.
All physical experiments were done by human scientists. They included protein expression, biochemical and structural analysis, and confirmation that the array produces short RNAs. The lab was set up in spring 2026, operates only at biosafety levels 1 and 2, and does not work with human pathogens. Researchers used Claude Code and Claude Science, the research product Anthropic launched in June.
The Caveats
Anthropic and outside observers have been clear about what the result does not show.
- The function is unproven. The preprint has not demonstrated that the enzyme acts on the RNAs the array produces.
- It was not completely unknown. Related reverse transcriptases had been seen before in jumbo phages. CEO Dario Amodei said a Stanford team previously described a system that is similar in some respects.
- Headlines may be ahead of the data. Some coverage said ART can "cut, copy and paste" DNA. More careful reports note that this describes other systems with a similar signature and has not been shown for ART.
- Other groups are doing similar work. AI-driven biology research is also under way at Stanford, UCSF and Google.
Amodei has called the finding preliminary. He has said it could turn out to be a new gene-editing mechanism with possible uses in gene therapy, depending on further experiments.
Why It Matters
Frontier labs have spent the past year claiming their models can make genuine scientific discoveries. This announcement is more checkable than most. It names a specific biological object, publishes the pipeline behind it, and discloses the compute used. That makes it a useful reference for what agentic research can do today: triage at a scale no human team could manage, with humans still responsible for experimental proof.
The economics are notable too. About 210 million tokens is a modest cost compared with a months-long bioinformatics project. If other labs can reproduce this kind of screening, the constraint in early-stage discovery moves from finding candidates to validating them in the lab.
The result also sharpens the biosecurity question Anthropic has raised itself. The company stresses that Claude does not run experiments. Amodei has said Claude might one day do lab work safely, but not yet. The more capable these agentic search pipelines become, the more their outputs, and not only their lab protocols, will need oversight.
What to Watch
The next milestones are peer review of the preprint, independent replication of the ART findings, and evidence of what the enzyme does to the short RNAs. If ART proves programmable, it would be the most significant AI-led biology discovery so far. If it doesn't, the 21-hour pipeline is still a notable demonstration of large-scale scientific triage by agents.
