Anthropic ran 950 Claude agents for 21 hours to find a new enzyme

Anthropic said in a post dated 23 September 2026 that Claude agents found a previously uncharacterised enzyme system, which it calls array-associated reverse transcriptase (ART), while searching a large database of DNA sequences. The company reports that roughly 950 agents ran for 21 hours and used 210 million tokens. Anthropic published the post to introduce a new life sciences research group and lab, so these are the vendor's own numbers about its own team.

The job, in numbers

• Scale: about 950 agents, 21 hours, 210 million tokens, according to Anthropic. By my arithmetic that is roughly 221,000 tokens per agent, and an average of about 45 agents running at any moment if they were spread evenly.
• Funnel: the agents gathered over 200,000 reverse transcriptases, picked out 3,500 new candidate systems, and narrowed those to the 20 most compelling, each analysed into a human-readable report.
• Time saved: Anthropic says that kind of analysis can take an expert scientist weeks to months. That is its estimate, not a measured comparison.
• Result: one agent noticed a repeating DNA pattern beside an unusual reverse transcriptase gene. The pattern and an accessory protein of unknown function define the new system.

How the workflow is built

Anthropic describes a repeatable pattern. Claude first reads the literature and reproduces established results from public data, as a check on its own methods. It then searches for family members or neighbouring genes that fit no described system, and writes a short report per candidate that proposes a function and sets out the evidence. A second pass critically evaluates that evidence. Anthropic says most candidates are eliminated at this stage, and that a survey may end with a single candidate worth testing or with none.

The team works in Claude Science and Claude Code, the same tools Anthropic says are available to any scientist, and sometimes with its own harness that coordinates many Claude sessions in parallel. Because a campaign yields hundreds to thousands of candidate reports, Anthropic says it studies which proposals its scientists judge worth testing and feeds that back into the instructions it gives Claude.

Where people stayed in the loop

Anthropic says its involvement was limited to the initial prompt and the lab work. All laboratory work is done by human scientists, at the lower biosafety levels (BSL-1 and BSL-2), and the company says it does not handle pathogens that can infect humans. It says AI-assisted lab automation is less suited to its ad hoc workflows. Claude helped interpret the data, and humans reviewed each report before anything was tested.

What the post leaves out

• Cost in money. Tokens are given; the price paid, and the compute behind any agents that found nothing, are not.
• Hit rate. We are told that most of the 20 reports were cut and that one system survived, but not how many reports reviewers rejected or why.
• Function. Anthropic says it does not yet know what ART does. It reports early experiments showing the repeat array is expressed as short RNAs, which it says suggests something analogous to CRISPR may be at play.
• Independent confirmation. The work is in a preprint. Feng Zhang, a CRISPR pioneer at MIT and the Broad Institute, is quoted in the post as saying the finding "is genuinely intriguing and merits further investigation"; that is a comment on the pattern, not a confirmation of what it does.

What a team can take from it

The useful part for anyone running agents on a large corpus is the funnel. Wide, cheap search came first, then a hard cut to a handful of reports a human could read, then an evidence review that discarded most of them, then expensive human verification only for survivors. Expect the same ratio in your own work: most of the agent output should be thrown away before a person spends time on it. Reproducing known results first is also a cheap test that your agent setup works before you trust what it finds.