Big tech is rapidly pivoting away from polite chat-bots and toward something with actual balance-sheet weight: patents, proprietary molecules, and automated R&D. According to Anthropic, its Claude models have autonomously uncovered a novel enzyme system that shares structural logic with CRISPR. This marks the inaugural output from Anthropic's newly minted wet lab—a facility designed less for academic curiosity and more for demonstrating that language models can actually run industrial science.
To execute the search, Anthropic fed massive genomic databases into a computational screening pipeline. As the company explained, human researchers did little more than write the initial prompt and handle the physical bench work, delegating the heavy pattern recognition entirely to the automated system.
Screening dynamics across agent swarms
Instead of traditional, painstakingly slow manual sequence inspection, the exploration relied on parallelized agent swarms. Over the course of 21 hours, roughly 950 Claude agents churned through 210 million tokens before flagging a peculiar repeating pattern in the dataset for human intervention.
"a previously uncharacterized enzyme system found in bacteriophages"
Subsequent laboratory checks confirmed the presence of a previously uncharacterized biological mechanism inside bacteriophages—viruses that hunt bacteria. It took nearly a thousand autonomous agents burning through compute and 210 million tokens just to surface a single anomaly.
Unclear utility and expanding laboratory scope
Despite the heavy computational overhead, the practical utility of this newly found mechanism remains entirely unproven. Anthropic admits it is still deciphering what the enzyme actually does, choosing to push the finding into public view early primarily to showcase Claude's capabilities and pitch its expanding ambitions in drug discovery. At this stage, calling this the next CRISPR is classic tech-PR overreach; we are looking at an expensive data-mining trick whose commercial viability is still an open question.
Yet the strategic signal is unmistakable. Big tech is moving past text-generation novelties and attempting to turn fundamental scientific discovery into a predictable machine pipeline. If agent swarms can compress months of genomic screening into a single day—even with unknown downstream utility—the economics of pharma and biotech R&D are in for a brutal, automated restructuring. The race is no longer about who writes the best essay, but whose silicon workforce can patent the next biology.