Leadership1 publisher2 min readPublished
Anthropic's new biology lab and 950 Claude agents together found a new enzyme system in 21 hours
The company's new life sciences group says agents working from a single prompt flagged an enzyme system with CRISPR-like DNA repeats, and its scientists are still running experiments to find out what the system actually does.
The Board Room · Leadership desk
What happened
- Anthropic has set up a life sciences research group and its own laboratory, formed in the spring of 2026, with one team covering both the training of Claude in biology and the bench experiments.
- The company says its involvement was limited to the initial prompt and the lab work, with Claude agents combing a DNA sequence database and using their own judgement to pick candidates.
- One agent flagged a repeating DNA pattern next to the gene for an odd-looking reverse transcriptase, and Anthropic calls the system array-associated reverse transcriptases, or ART.
- The reverse transcriptase had been identified in earlier studies, and Anthropic says Claude appears to be the first to notice the repeat array beside it plus an accessory protein of unknown function.
Compiled by The Board RoomSomething wrong?How this is made
Why it matters
- constraint The search was the fast part, and functional characterization sets the pipeline's throughput: buying more agent time lengthens the candidate queue without increasing the number of answers a lab can return.
- decision A research leader pricing this has to decide whether a newly noticed co-occurrence of known parts counts as discovery, because that judgement determines whether agent search is funded as a screening tool or as a hypothesis generator.
- exposure Until an outside group reproduces it, anyone citing ART is relying on a pre-print from the company that sells the model which produced it.
- precedent A model vendor that owns bench capacity can validate its own outputs before publishing, and other vendors now have a template for making discovery claims without a partner institution.
Roughly 950 agents used 210 million tokens on that search, which works out to about 221,000 tokens each [5][15]. One long context window per agent. The run was broad and shallow: hundreds of parallel passes over the same sequence database, all of them working from one instruction to look for interesting new examples of reverse transcriptases [16].
Anthropic says the combination of characteristics in ART has only ever been found together in a handful of other systems, all of which are programmable and perform operations like cutting, copying and pasting DNA [9]. The company says it does not yet know what ART does, that the work to establish its primary function is ongoing, and that it has released a pre-print [8][14][13].
Feng Zhang, a pioneer of CRISPR genome editing and a professor at MIT and the Broad Institute, reviewed the pre-print [11]. "The identification of RNA-repeat arrays associated with reverse transcriptases is genuinely intriguing and merits further investigation," he said [10]. Anthropic published the result itself, and says it shared early "both to demonstrate Claude's capabilities and to give the broader community insight into what we're working on" [12].
The 21 hours covered the search only [5]. The prompt and the bench work came from the humans in this account [4].
The enzyme itself was already described in the literature, and the pairing beside it had not been reported [7]. The one outside scientist on the record called it worth investigating [10]. The assay settles it, and it has not been run to conclusion [14].
The decision this quarter is a budgeting one. Anthropic's post does not address authorship, credit or patent rights [18]. It does show where the human labour sat. If one prompt and a day of agent time produce a queue of candidates, then candidate ranking and wet-lab throughput decide how many of them ever get an answer. Anthropic's own response to that was structural, building the lab and putting a single team on both training Claude in biology and running the experiments [1].
The ownership-of-the-hypothesis question is real, and it will be settled by journals, patent offices and institutional policy over years, not by one pre-print. Anthropic says ARTs are found in bacteriophages [6] and that it still has to determine what they do [14].
What to watch
- Whether Anthropic's lab work establishes ART's primary function, and whether it cuts, copies or pastes DNA the way the comparable systems do.
- Whether a group outside Anthropic reproduces the finding, and whether the pre-print clears peer review.
- Whether Anthropic states how authorship and inventorship are assigned when an agent, not a scientist, generates the candidate.