Anthropic announced research on September 23 in which a large group of Claude agents helped researchers locate a biological system that had not previously been characterized. The company calls the system ART. According to Anthropic, the work involved broad searches of biological databases, screening thousands of candidates and subsequent laboratory testing performed by human researchers.
The announcement is interesting not just because of the finding itself, but because of how it was found. Anthropic says approximately 950 Claude agents operated for about 21 hours, consumed approximately 210 million tokens and scanned information on more than 200 thousand reverse transcriptases. The process narrowed the list to approximately 3,500 candidates and produced 20 reports submitted for human review.
What is the ART system?
According to Anthropic's description, ART includes a reverse transcriptase, a nearby gene and an array of repeated DNA sequences. Initial experiments found that the array produces short RNA molecules. At this stage, however, the system's main biological role remains unknown.
This matters because some discussion of the announcement focuses on a certain resemblance to CRISPR. The resemblance Anthropic describes concerns the structure of repeated sequences and the genetic organization around them. There is currently no proof that ART operates as a gene-editing system, is programmable or can serve as a practical tool for modifying DNA.
Claude did not conduct the research alone
Anthropic presents the project as a demonstration of using large numbers of agents for scientific search and analysis at a scale difficult to achieve manually. However, the process was not autonomous research from end to end.
The agents scanned databases, analyzed candidates and produced reports. Human researchers selected promising results, designed further testing and conducted laboratory experiments. In other words, Claude's contribution was significant according to the company, but the finding was not achieved without human screening, judgment and verification.
Reuters reported that the reverse transcriptase at the center of the system had already appeared in previous studies. The novelty claim is therefore not that the enzyme itself had never been observed, but that the team identified the broader system and the connection between the enzyme, nearby gene and repeated DNA sequences.
What the research does show
If the findings withstand further testing, the work could be a useful example of how AI agents may help scientists scan enormous databases and seek connections difficult to identify manually. In this case, the value is not that the model "did biology" alone, but that it narrowed a large search space and suggested directions for laboratory testing.
This is a type of work where AI systems can be particularly efficient: repetitive searching, comparing large numbers of sequences, gathering clues from different sources and generating hypotheses a researcher can test.
What remains unknown
The study was published as a preprint and has not yet undergone peer review. ART's role is unknown, no proof has been published that it can be programmed, and no other group has yet independently replicated the findings.
It is also difficult to quantify precisely how much of the result came from Claude's reasoning capability, how much from the search and automation tools incorporated into the process, and how much from the human team's task selection, screening and decisions.
Anthropic's announcement is therefore mainly an interesting indication of the direction agent-based scientific research may take. It is not yet proof that AI models can conduct independent biological research or that ART is a new gene-editing tool.
Why the term "discovery" requires care
In scientific research, identifying a computational pattern is only one stage. In ART's case, the agents proposed connections and produced reports, but human researchers were needed to select candidates and test them in the laboratory. Finding the system within existing databases is also not the same as creating a new biological entity.
The achievement, if verified, is therefore mainly the identification of a biological organization not previously characterized as a complete system. The next step will be to understand what it does, whether it appears in other contexts and whether the findings can be independently replicated.