AI Agents Handle Lab Work, Humans Still Drive Discovery

Anthropic told the world last Wednesday that it had opened a molecular biology lab earlier this year, one staffed by 950 Claude agents instead of postdocs. The pitch: let AI read through millions of cataloged DNA sequences, flag patterns worth investigating, and hand the promising ones to human scientists for actual lab testing. After 21 hours of work, the company says its system turned up something new, a repeating pattern surrounding a known enzyme that hadn't been documented before.

Anthropic didn't stop at describing the find. Its announcement compared the pattern to the research that led to CRISPR, the gene-editing tool that reshaped medicine and biology over the past decade. That comparison is doing a lot of work.

CRISPR came from years of grinding research into bacterial immune systems before anyone understood its potential. Calling an unindexed sequence pattern "reminiscent" of that breakthrough sets a bar the finding hasn't cleared yet, and it's the kind of language that decides whether the public sees this as a genuine discovery or as a company marketing its own tools. That gap, between what an AI system actually did and what a press release claims it did, is where the argument over scientific credit is happening now.