950 Claude agents, 21 hours: fan-out is the real workload

The interesting number in Anthropic's enzyme announcement is not the enzyme. It is 950. On September 23, Anthropic published that Claude identified a previously uncharacterized enzyme system it calls ART — array-associated reverse transcriptases — in bacteriophages. The work took roughly 950 Claude agents, 21 hours and 210 million tokens, scanning over 200,000 reverse transcriptases down to 3,500 candidate systems and then 20 finalists. That shape — massive fan-out, brutal filtering, a human at the end — is the parallel agents pattern that actually earns its token bill.

What actually happened

Anthropic's life sciences research group, formed in spring 2026, pointed a swarm of agents at a DNA sequence database. The system they surfaced has three parts: a reverse transcriptase, a partner gene next to it, and a long array of evenly spaced DNA repeats. The repeat array is what makes it look CRISPR-adjacent, and the array expresses as distinct short RNAs. The underlying reverse transcriptase had appeared in prior work; the defining combination had not been described.

Two details deserve as much attention as the headline.

First, Anthropic says plainly: "we don't yet know its function." Feng Zhang of MIT and the Broad Institute called it intriguing and worth further investigation — which is a careful scientist's way of saying this is a lead, not a result.

Second, all the lab work was done by human scientists. Expressing the protein, characterizing it biochemically and structurally — that happened at a Bay Area bench at BSL-1 and BSL-2, by people. The agents did search and triage. Nothing was discovered by a model in the sense a headline implies; a very large haystack got sorted very fast.

Why parallel agents matter for your business

You do not have a DNA database. You do have haystacks, and this is the template for them.

The winning shape is search and triage, not judgment. 200,000 → 3,500 → 20 → humans. Every step narrows a corpus too large for a person while a person still makes the final call. The equivalents in an operating business are mundane and real: 4,000 supplier invoices checked against contract terms, every product listing compared to the spec sheet it came from, two years of support tickets sorted for the twelve that describe the same undiagnosed bug, a competitor's full catalog priced against yours.

Fan-out beats depth for this class of work. One agent thinking hard about one invoice is a chatbot. Nine hundred cheap agents each doing one narrow pass, with results collapsing into a ranked shortlist, is an instrument. The prompt engineering that matters is the filter criteria, not the persona.

Budget the run, not the month. 210 million tokens sounds enormous until you scope it as one job with a beginning and an end. Price the run against what the alternative costs — the contractor you would hire, or the review that simply never happens because nobody has three weeks. Then decide. That is a purchase decision, not an AI strategy. Run your own numbers before you believe anyone's.

Publish your uncertainty the way they did. Anthropic shipped a finding with "we don't know what it does" attached. When your agent flags 20 invoices as suspect, that is 20 candidates for a human, not 20 findings. Automations that state confidence get trusted and kept. Automations that assert get switched off after the first embarrassing miss.

Sources: Anthropic.