Anthropic says 950 Claude agents found an enzyme system. A CRISPR expert objects
Anthropic says its agents scanned over 200,000 reverse transcriptases and found a system it appears to be first to describe. Mammoth Biosciences co-founder Lucas Harrington says the search method is decades old and the hard part, finding what it does, is still missing.
210 million tokens, 950 agents, one shortlist of 20. We did the per-agent math on Anthropic's enzyme claim.
Anthropic said on September 23, 2026 that about 950 Claude agents, running for 21 hours and using 210 million tokens, found a family of genes it calls ART. A day later, Lucas Harrington, a co-founder of Mammoth Biosciences, said on X that the search method is decades old and the hard part has not been done, The Decoder reported.
Anthropic's post says Claude discovered the system autonomously. It also says human involvement was limited to the first prompt and the lab work, and that the agents picked candidates using their own judgment.
From 200,000 sequences to 20
According to Anthropic, the agents gathered more than 200,000 reverse transcriptases, worked through the distinct families, and flagged 3,500 candidate systems. The 20 most promising got detailed reports. The work ran in Claude Science and Claude Code, on what the post calls a harness of Anthropic's own that coordinates many Claude sessions in parallel. Human scientists did all of the lab work. The post does not name the model.
The biology, briefly. ART stands for array-associated reverse transcriptase: a reverse transcriptase, a partner gene beside it, and a long array of evenly spaced DNA repeats, found mostly in bacteriophages, the viruses that infect bacteria. Lab work so far shows the array is expressed as a set of short RNAs. Anthropic says it does not yet know what the system does. The results sit in a preprint hosted on Anthropic's own site, not a peer-reviewed journal paper. Feng Zhang, a professor at MIT and the Broad Institute, reviewed the preprint, and Anthropic quotes him calling it an exciting example of agents contributing to biological discovery.
Known since 2008, says Harrington
Harrington, who did his PhD under Jennifer Doudna and co-founded Mammoth with her, says genome mining has existed for decades and that similar systems have been known since 2008. "The hard part is figuring out what a system actually does, and Anthropic hasn't shown that," he says, as quoted by The Decoder. Mammoth uses genome mining to hunt for CRISPR systems itself, a point The Decoder notes. The Decoder reports no response from Anthropic.
Anthropic's post also says its work on what ART does is ongoing. The open question is whether finding and describing the system counts as a discovery.
Why a build studio cares
Run the published numbers. 210,000,000 tokens across 950 agents is about 221,000 tokens per agent. Across 200,000 sequences it is an average of about 1,050 tokens each. So 950 is a throughput figure. It shows a harness can keep that many sessions busy for 21 hours, and it proves nothing about judgment. The judgment claim lives in the cut from 3,500 candidates to 20, and that is where we would look first on any agent pipeline sold as autonomous: what the first prompt said, who checked the cut, and what a plain mining script with no agents finds on the same data. Anthropic's post gives neither a baseline nor a cost. We build with Claude Code every day, and we have no stake in this result.
Next step: read Anthropic's post, which links the preprint, and The Decoder's report on Harrington's objection. If a vendor is pitching you an agent pipeline on a headline number like 950 agents, write to us at hello@gattyworks.com.