Claude agents found a CRISPR-like enzyme system

Anthropic announced on 2026-09-23 that Claude agents autonomously discovered a novel enzyme system associated with an array of DNA repeats, a pattern reminiscent of CRISPR. The company described the finding in a post on anthropic.com and released a pre-print. The work comes from a life sciences research group and laboratory Anthropic formed in spring 2026 to test whether general AI models can systematize and accelerate biological discovery.

What the agents actually did

Anthropic says it gave Claude a prompt to search a large DNA sequence database for interesting new examples of reverse transcriptases (RTs), enzymes that copy RNA into DNA. The company states its involvement was limited to the initial prompt and the lab work. According to the post, roughly 950 agents spent 21 hours searching, using 210 million tokens. They gathered over 200,000 RTs, picked out 3,500 new candidate systems, and narrowed those to the 20 most compelling, each written up as a human-readable report. One agent noticed a repeating pattern of DNA sequences next to the gene for an odd-looking RT.

After further analysis and lab testing, Anthropic recognized the pattern as a previously uncharacterized enzyme system found in bacteriophages, which it calls array-associated reverse transcriptases (ART). The underlying RT had been identified in previous studies. Anthropic says Claude appears to be the first to notice the system's defining features: an associated array of non-coding DNA sequences and an additional accessory protein of unknown function. Feng Zhang of MIT and the Broad Institute called the identification of RNA-repeat arrays associated with reverse transcriptases genuinely intriguing and said it merits further investigation.

What is not yet known

Anthropic states that work to understand the primary function of ARTs is ongoing. The system's three parts are described, and first experiments show the ART array is expressed as a set of distinct short RNAs, but how ART works is not established. The company says it shared the results early to demonstrate Claude's capabilities and to give the community insight into the work.

For readers tracking AI in research, the practical point is narrower than the headline. This is a vendor account of one research program, with a pre-print and a named outside comment, not an independent evaluation of agent reliability. The reported scale — hundreds of agents, tens of millions of tokens, one candidate worth reporting — is also a useful signal about the ratio of hypotheses generated to candidates that survive review. Anthropic says most candidates are typically eliminated in follow-up analysis, and a survey may end with a single candidate or none.

Teams considering similar agent-driven literature or database sweeps should treat the output as a queue of hypotheses requiring expert review and laboratory confirmation, not as findings. The ART result is a candidate for further study, not a validated tool.