Anthropic Says Claude Agents Found a New CRISPR-Like Enzyme System — and Nobody Knows What It Does Yet

Anthropic has announced a life sciences research group and a wet lab of its own, and led the introduction with a result: the company says Claude agents, working from a single high-level prompt, spotted a previously uncharacterized enzyme system in bacteriophage genomes that carries a repeat array reminiscent of CRISPR. Anthropic calls the system ART, for array-associated reverse transcriptases, and says its function is still unknown. The findings are described in a preprint and a technical report; they have not, as far as the announcement states, been peer reviewed.

What Anthropic says happened

According to Anthropic, the team gave Claude a prompt to search a large DNA sequence database for interesting new examples of reverse transcriptases — enzymes that copy RNA into DNA. From there, the company says, human involvement was limited to that initial prompt and the subsequent lab work. Roughly 950 agents ran for 21 hours and consumed 210 million tokens. Anthropic says the agents assembled more than 200,000 reverse transcriptases, identified about 3,500 candidate systems, and narrowed those to 20 that they wrote up as human-readable reports.

One agent, per the post, noticed a tandem repeat array sitting next to an unusual-looking reverse transcriptase gene and flagged it in its own output as a possible CRISPR-like repeat array. Anthropic says the agent then counted the repeats, measured spacing, compared the arrangement to known reverse transcriptase systems, searched the literature for prior reports, and filed a report for human review.

The resulting system, as described, has three parts: the reverse transcriptase, an adjacent partner gene of unknown function, and a long array of evenly spaced DNA repeats. Anthropic says the underlying reverse transcriptase — found in a jumbo phage — had already been identified in earlier studies, but that Claude appears to be the first to flag the array and the accessory protein as defining features of a system. Early experiments in Anthropic's lab, the company says, show the array is expressed as a set of distinct short RNAs, which it suggests may be loosely analogous to how CRISPR arrays supply the guide RNAs that make Cas systems programmable.

Anthropic is careful on the central point: it does not know what ART does. The company frames the characteristics as a combination found together in only a handful of other systems, all of which are programmable and perform cut-copy-paste operations on DNA. That is a statement about family resemblance, not about demonstrated function.

The outside comment, and what it actually says

Anthropic quotes Feng Zhang of MIT and the Broad Institute, one of the pioneers of CRISPR genome editing, who reviewed the preprint. His comment calls the identification of RNA-repeat arrays associated with reverse transcriptases intriguing and says it merits further investigation, and encourages other scientists to explore how AI can support research.

That is a meaningful endorsement of interestingness, not of function or utility. Anyone reading the announcement as validation that ART will become a gene-editing tool is reading past what the quote says. Anthropic did not disclose whether other outside researchers reviewed the work, or whether the preprint has been submitted to a journal.

The workflow is the more portable claim

The enzyme may or may not matter. The described process is the part product teams should look at closely, because it is a concrete template for agentic research pipelines in any data-rich domain.

Anthropic describes a repeating pattern: Claude surveys a protein family, reads the relevant literature, reproduces established results from public data as a self-check on its methods, then searches for family members or genomic neighbors that fit no described system. For each candidate it writes a short report proposing a function and laying out supporting evidence. A follow-up pass has Claude critically evaluate that evidence, and the company says most candidates die at this stage. A survey may end with one candidate worth testing — or none.

Two details stand out. First, the reproduce-known-results step is a cheap, legible way to catch a tool-use pipeline that is silently broken, and it transfers directly to non-biology agent systems. Second, Anthropic says it has started treating the hypotheses themselves as an object of study — comparing which proposals human scientists chose to test against those they discarded, then folding that signal back into the instructions. That is taste elicitation, and it is arguably the harder half of the problem: an agent fleet that generates thousands of plausible reports is only useful if something filters them at a rate humans can absorb.

Anthropic says the work runs in Claude Science and Claude Code, the same products available to other users, sometimes with an internal harness coordinating parallel sessions. The harness is not described in detail and does not appear to be available.

What the announcement leaves out

The honest reading is that this is one publicized success from an internal program formed in spring 2026. The post does not say how many campaigns were run, how many produced nothing, or what the total compute spend has been — the 210 million tokens covers this single search, not the program. Without a denominator, the hit rate is unknowable.

There is also no baseline. Anthropic notes that this kind of analysis can take an expert scientist weeks to months, but the post does not compare the agent pipeline against a conventional genome-mining workflow or against non-agentic computational tools on the same database. "Twenty-one hours versus weeks" is a framing, not a controlled comparison, and the human time spent building the harness, curating the prompt and doing the lab validation is not accounted for.

The word "autonomously" is doing work, too. Humans chose the domain, wrote the prompt, reviewed the report, decided what to test, and ran every bench experiment. Anthropic says all lab work is performed by human scientists, and that its experiments with automating lab work through the Model Hardware Standard were a poor fit for the ad hoc workflows of molecular biology. That is a useful admission about the limits of current lab automation.

The safety framing

Anthropic states the Bay Area lab operates only at BSL-1 and BSL-2 and does not handle pathogens that can infect humans. That disclosure lands in the same week's context as the company's Life Sciences Verification Program, which it describes as giving verified life science professionals access to models with safeguards that are more permissive for biology work.

Those two things sit in tension worth naming. A frontier lab is simultaneously loosening biology-related refusals for vetted users and demonstrating that its agents can autonomously identify uncharacterized systems in phage genomes. Anthropic has not published, in this post, how the verification program's gating maps to the kinds of capability shown here, or what review the internal research group's outputs pass through before release.

What to watch

• Whether follow-up experiments establish an actual function for ART, or whether the CRISPR resemblance turns out to be structural coincidence. Short RNAs expressed from an array is suggestive; it is not programmability.
• Whether independent labs can reproduce the finding from the preprint and the public sequence data, and whether the work clears peer review.
• Whether Anthropic publishes failure statistics — campaigns run, candidates tested, candidates that did not pan out. That number determines whether this is a repeatable method or a well-marketed lottery ticket.
• Whether the parallel-session harness, or the taste-elicitation loop behind it, ships in any form outside Anthropic. The described workflow is the transferable asset here.
• How the Life Sciences Verification Program's relaxed biology safeguards are audited in practice, given that agentic genome mining is now a demonstrated capability rather than a hypothetical one.