What Anthropic says Claude found
On 23 September, Anthropic announced that Claude had found a previously unknown enzyme system in bacteriophage DNA. It calls the system array-associated reverse transcriptases, or ART: a reverse transcriptase, a partner gene, and an array of DNA repeats. The repeats earned the CRISPR comparison, and the comparison moved the story out of the science pages. Shares of gene-editing companies including CRISPR Therapeutics, Intellia Therapeutics and Beam Therapeutics dipped on the news.
The biology is interesting, but most readers of this blog can't use it. What they can use is how the search ran, and how often it would find ART again. According to The Next Web's reading of the preprint, Anthropic reran the same campaign ten more times and none of the reruns found the array. That is one hit in eleven runs. If you build agents that search, triage or investigate, that ratio deserves more attention than the enzyme.
The pipeline: one brief, 949 sessions, 19 reports
Anthropic's post gives the scale in round numbers: roughly 950 agents searching for 21 hours and using 210 million tokens. They gathered over 200,000 reverse transcriptases, picked out 3,500 new candidate systems and narrowed those to 20 for deeper analysis.
The preprint figures, as reported by The Next Web, are more precise. Agents running Claude Mythos 5 searched a database of 1.9 billion protein clusters. The campaign ran for 21.5 hours without human input, across 949 agent sessions and 215.6 million tokens. It scored 3,564 candidate partner families and filed 19 reports for human review.
Read as a systems diagram, that is a funnel:
- Wide fan-out: hundreds of parallel sessions working one brief over a very large search space.
- Automated narrowing: hundreds of thousands of enzymes cut down to a few thousand scored candidates.
- Human gate at the end: a couple of dozen written reports for people to read, followed by lab work done by human scientists in Anthropic's Bay Area lab. Anthropic says that lab works only at biosafety levels BSL-1 and BSL-2 and does not handle human pathogens.
The tooling is not exotic either. Anthropic writes: "We do our work in Claude Science and Claude Code, the same tools available to any scientist, and sometimes with a harness of our own that coordinates many Claude sessions running in parallel."
So the ingredients are an off-the-shelf agent, a coordinator and a lot of tokens. Engineering teams use the same shape for code migration, log triage and security review, so the lessons carry over.
The number the headlines skipped: one in eleven
Two details in The Next Web's account of the preprint matter more to engineers than anything about repeats or reverse transcriptases.
The first is the rerun result. The same campaign, run ten more times, never surfaced the array again. The finding was solid enough to take into a lab, but the pipeline that produced it did not produce it reliably. If Anthropic had run the campaign only once and landed on one of the other ten outcomes, there would have been nothing to announce.
The second is what tools cost. In fixed tests where the model was given the DNA directly, it described the array in at least 90% of attempts. When it had to work through files and tools to reach the same material, the rate fell as low as 32%. The model could recognise the pattern. The agent often never got to the point where it needed to.
That second finding is the one to pin to the wall. Most agent evaluations test the model's judgement on inputs placed in front of it. Production agents have to find those inputs first: open the right file, run the right query, page through the right results, decide what to ignore. Each of those steps is a chance to wander off. A capability that shows up 90% of the time can become a one-in-three outcome once navigation is part of the job. A single successful run won't tell you which of those numbers you are living with.
Together, the two findings say that long-horizon agent search behaves less like a function and more like a sampler. Each run is a draw. Some draws walk into the interesting part of the space; most do not. A single run tells you what one path found, not what the space contains.
What the scientists are actually arguing about
Most of the public argument has been about whether ART is "the next CRISPR". Anthropic is careful here: it says the function of ART is not yet known. The Next Web reports that the preprint says enzyme activity has not yet been shown. The preprint has not been peer reviewed.
Outside reactions split along predictable lines. Feng Zhang of MIT and the Broad Institute, quoted by Anthropic, said: "The identification of RNA-repeat arrays associated with reverse transcriptases is genuinely intriguing and merits further investigation." Stanley Qi of Stanford, speaking to Al Jazeera, pointed at the method rather than the molecule: "What stands out is its ability to recognize an unusual biological pattern that was difficult to detect before, and to pursue it comprehensively as a research question."
Kevin Blake of Washington University School of Medicine, also speaking to Al Jazeera, pushed back on the framing: "There's nothing to indicate this is a rival to CRISPR-the-technology, or could be developed into any kind of therapeutic or practical application."
These positions don't really contradict each other. Blake is talking about applications, and nobody has shown one. Qi and Zhang are talking about the finding and the search that produced it. Engineers should keep the two claims apart too. "The agent found something worth a lab's time" is a claim with evidence behind it. "The agent found a new gene-editing tool" is a claim the evidence does not yet support. The dip in gene-editing shares says more about how headlines travel than about the science.
What to do if you run agents for search or discovery
If your agents look for things like bugs, anomalies, fraud patterns or relevant documents, the ART campaign shows where the risk sits. The question is less whether the model is capable enough and more whether a given run gets there.
Treat run-to-run variance as the main risk. Before you trust a pipeline, run it on the same brief several times and compare what each run surfaces. If the runs have little in common, a single run is a sample, not a survey.
Measure recall across reruns, not per run. Plant known targets in your search space, such as a seeded bug, a known-bad transaction or a document you know is relevant. Then count how many runs find each one. That gives you a real recall figure and tells you how many runs you need for acceptable coverage.
Test through the tools, not around them. The drop from over 90% to as low as 32% is the gap between testing the model and testing the agent. If your evals hand the model the relevant input directly, they measure the easy part. Build at least some evals where the agent has to find the input with the same files and tools it gets in production.
Budget for repeated or ensembled runs. If one run finds a target a third of the time, several independent runs with pooled results give you much better odds. The runs have to be genuinely independent, so vary seeds, starting points or partitions of the search space. Price this in up front. A single campaign of 215.6 million tokens is exactly the kind of bill that tempts teams to skip the reruns.
Keep a human-reviewed funnel at the end. Anthropic's campaign ended in 19 reports and a lab, not in an automatic claim. Whatever your domain, the agent's job is to narrow the space to something a person can check. Keep the output small and the reports specific, and make sure someone reads them.
Report the denominator. When an agent finds something, say how many runs it took and how many missed it. The preprint's rerun section does exactly that, and it is why the one-in-eleven figure is public at all. Your internal write-ups should do the same.
The enzyme may or may not turn out to matter. The one-in-eleven figure already matters to anyone who puts agents on open-ended search.
Comments (0)
No comments yet. Start the conversation.