Claude found an unusual DNA arrangement. ART is not a CRISPR tool yet

What the agents found

Anthropic asked a multi-agent Claude Code harness to look for novel reverse-transcriptase systems. Workers built sequence profiles, searched 1.9 billion protein clusters, classified roughly 200,000 reverse-transcriptase clusters, and investigated unusual neighboring genes. Supervising agents could open follow-up tasks, while a shared record kept plans and results available across the campaign.

One follow-up worker inspected DNA beside a reverse transcriptase from jumbo phages and noticed a tandem repeat array. The broader ART pattern combines repeat-rich non-coding DNA, a reverse transcriptase, and a nearby partner gene. The enzyme family itself was not newly created or proven by Claude; the reported discovery is the previously undescribed genomic arrangement around it.

What the wet-lab work confirmed

The researchers found 95 distinct ART reverse-transcriptase clusters, including 28 with a detectable upstream repeat array. The arrays differ from CRISPR arrays in spacer length and organization, and the team found no nearby cas genes.

Public RNA-sequencing data showed that one ART array was highly expressed during Staphylococcus phage SA1 infection. The team also expressed the SA1 system in E. coli and observed discrete short RNAs from the array. This supports the claim that the array is transcribed and processed. It does not establish the system's biological function.

Why the reruns matter

The original campaign used 119 tasks, 949 agent sessions, 215.6 million tokens, and 21.5 hours of wall-clock time. The authors then ran the same campaign ten more times. ART loci appeared in most completed censuses and two workers investigated the lineage, but none read the upstream DNA and none of the ten reruns rediscovered the array.

A fixed-input benchmark produced a more repeatable result. When capable models received ART loci directly in context, they recognized the repeat array in at least 90% of attempts. Tool-rich runs could do worse because agents sometimes never read enough contiguous DNA. That is a useful agent-design lesson: access to a file does not prove the model inspected the decisive evidence.

What ART is, and what it is not

• ART is a candidate biological system built around array-associated reverse transcriptases, repeat arrays, and partner genes.
• The repeat architecture is reminiscent of CRISPR, but ART is not a demonstrated CRISPR system or gene-editing tool.
• Its biological function remains unknown. The proposed defense mechanism and programmable behavior are hypotheses for future experiments.
• The September 23 preprint has not been peer reviewed, and all named authors are Anthropic researchers.
• The work shows a promising AI-assisted discovery workflow, not a general proof that an agent can reliably make autonomous scientific discoveries.

A better pattern for AI-assisted discovery

Use agents to widen a search, preserve complete transcripts, and rank anomalies for expert review. Then separate computational observations, reused public data, new laboratory measurements, mechanistic hypotheses, and independently reproduced findings. Each step supports a different strength of claim.

For reproducibility, report the model and harness, search space, task graph, token and time budget, selection criteria, negative results, and rerun rate. Compare the agent workflow with a strong conventional bioinformatics baseline. The ART preprint gives unusually useful operational detail, but it does not include that baseline comparison.

How to read the announcement

Anthropic's announcement is the dated company account. The accompanying preprint supplies methods, measurements, reruns, and author affiliations. Hacker News discussion surfaced two fair cautions: the reverse transcriptase was already known, and the phrase CRISPR-like can suggest an application that the experiments have not established.

The fair short version is that an agent campaign noticed a previously undescribed genomic arrangement, and human researchers confirmed that its repeat arrays produce short RNAs. Whether ART performs a useful or programmable operation is still an open scientific question.