Anthropic's agents found a new phage enzyme system once in eleven tries

Anthropic ran the same campaign ten more times. No rerun even read the DNA upstream of the enzyme, so all ten missed the array ^[14]. One success in eleven attempts is a 9 percent hit rate on the run that produced the headline ^[19].

That is not a knock on the finding. It is the shape of the thing anyone evaluating agents for discovery work has to plan around. The output was a side path an agent wandered down while doing something else ^[7].

Anthropic tested the same models in a narrower setup and the gap is instructive. Given the DNA directly, its four most capable models described the array in at least 90 percent of attempts ^[15]. Handed files and tools instead, the rate fell as low as 32 percent, often because the model never read enough DNA to see a full repeat ^[16]. The bottleneck was context management: whether the agent, left to manage its own context and tool calls, ever pulled the relevant bases into view.

The scoreboard from the 21.5 hour run: about 200,000 enzyme clusters recovered, 3,564 candidate partner families scored, 19 reports filed for a human to read ^[6]. Anthropic says genome mining of this kind could take an expert scientist weeks or months ^[17]. Nineteen reports is a manageable reading pile for a small team, which is the part of this that generalises to anyone building a screening workflow: the agents' job was to cut a 1.9 billion cluster database ^[5] down to something a person could actually review.

The biology is unresolved, and Anthropic says so. The enzyme itself was already known from earlier studies; what Claude appears to have spotted first was the repeat array and the partner protein beside it ^[8]. The arrays hold 3 to 21 copies of a short repeat and carry none of the cas genes that sit near CRISPR systems ^[13]. In published data from a Staphylococcus phage, the repeat-derived RNAs made up as much as 8 percent of the phage's RNA fifteen minutes after infection ^[3]. The preprint says the team has not shown that the enzyme is active, or that it works on those RNAs ^[4], and it has not been peer reviewed ^[2].

Feng Zhang of MIT and the Broad Institute reviewed the preprint, and Anthropic published his comment with its post ^[9]. "The identification of RNA-repeat arrays associated with reverse transcriptases is genuinely intriguing and merits further investigation," Zhang said ^[10].

For a lab head reading this on Monday, the forcing question is which of two things you are buying. One is a search that reliably narrows a huge database and hands you a short stack of reports; on Anthropic's own numbers that part worked, and the model-with-tools failure rate tells you how much redundancy to build in. The other is the specific novel find, and the ten failed reruns say you cannot schedule that. Anthropic's Bay Area lab works only at biosafety levels 1 and 2, handles no human pathogens, and human scientists do all the wet work ^[11].