Anthropic’s new life-sciences group reports that roughly 950 Claude agents, running for 21 hours and consuming 210 million tokens, searched more than 200,000 reverse transcriptases. They produced 3,500 candidate systems, narrowed those to 20, and identified a repeat array beside a reverse-transcriptase gene in a jumbo phage. Lab work confirmed that the array is expressed as distinct short RNAs. The team calls the system array-associated reverse transcriptases, or ARTs. Anthropic report and preprint, September 23
The resemblance to CRISPR is structural, not functional. Anthropic does not yet know what ART does, and the work is a company preprint rather than peer-reviewed evidence of a programmable tool. Still, this is a better agent-swarm demonstration than a synthetic benchmark: the system reproduced known results, searched a large sequence space, proposed candidates, rejected most of them and left a small set for wet-lab testing. The economic constraint has moved from generating hypotheses to choosing which ones deserve scarce experimental attention.