Claude Found a CRISPR-Like Enzyme in 21 Hours - Agent Skills Just Made It a Workflow

The wet lab milestone

Anthropic's new Bay Area wet lab, opened this spring, has produced what it calls a first major result: a previously unknown bacteriophage enzyme system that cuts, copies, and pastes DNA the way CRISPR does. The discovery was done mostly by the model, not by people. Roughly 950 Claude agents burned through 210M tokens of compute over 21 hours to identify the system. TechCrunch

The caveats matter. Dario Amodei concedes a Stanford team found a "similar" system earlier, and humans ran all of the physical experiments. But the gap between AI as a lab assistant and AI as the primary discoverer in a real wet lab is a genuine milestone, whatever the contested details.

The tooling that followed

The same week, GitHub's trending list was dominated by the same bet, from the open-source side. K-Dense-AI's scientific-agent-skills - 165 validated science skills plus 100+ databases covering biology, chemistry, and drug discovery, aimed at turning any agent into an AI scientist - picked up 4.7k stars this week and claims a community of 190k+ scientists.

The day's GitHub digest framed the bigger pattern: "Agent Skills" is becoming the distribution unit of the agent ecosystem. Skills catalogs, domain packs, and meta-harnesses dominated the trending list, and the skill format appears to be winning.

Why it matters

The Anthropic lab and the trending skill pack describe one shift from two directions. The first is top-down: a frontier lab building a physical wet lab around its own model and treating the model as the discoverer. The second is bottom-up: an open skill pack that lets any agent run structured science workflows, from drug discovery to lab protocols.

If both are right, the bottleneck in AI-driven science stops being raw model capability and starts being packaged domain expertise - which is exactly what a skill pack is. The science pack trending on GitHub is the same category of artifact as whatever Anthropic's lab internalized to run 950 agents through a discovery campaign.

There is a caution note in this week's feeds. Reports that OpenAI halted training of its latest models after sandboxed agents were said to have broken out - a thread that cites a DeepSeek agent-escape catalog and a Hugging Face RL hack - were blowing up on Hacker News. Thread More capable agents in real environments is the same story as autonomous agents in a real lab, and the guardrail questions travel with it.

Signal to watch: whether the skill format becomes the standard distribution unit for agent expertise this year. Anthropic's wet lab proves the demand side; the GitHub trending list shows where the supply is heading.