Anthropic just announced something that sounds like science fiction: a swarm of Claude agents, working mostly on their own, found a previously unknown biological system hiding in DNA data.
The humans wrote one prompt and ran the lab tests afterwards. Everything in between was the agents. Here's the plain-English version of what happened.
What Anthropic Actually Announced
Anthropic has quietly built its own life sciences research group — and its own wet lab in the Bay Area. The team's goal is to find out whether general-purpose AI models can speed up the kind of discovery that has historically happened by accident.
Their first big result: Claude found a new enzyme system that Anthropic is calling ART (array-associated reverse transcriptases). It lives mostly in bacteriophages — the viruses that infect bacteria.
Nobody knows yet exactly what ART does. But it shares a rare combination of traits with a small handful of other systems, and every one of those turned out to be programmable tools that can cut, copy or paste DNA. The most famous member of that club is CRISPR.
The Numbers Behind the Search
This wasn't one clever chatbot session. It was a coordinated campaign:
- ~950 Claude agents running in parallel
- 21 hours of searching
- ~210 million tokens consumed
- 200,000+ reverse transcriptases gathered from a massive DNA database
- 3,500 new candidate systems picked out
- 20 top candidates written up as human-readable reports
Anthropic says this kind of genome-mining work would normally take an expert scientist weeks to months.
The underlying enzyme had actually been spotted in earlier studies. What nobody had noticed was the stuff next to it: an array of repeating DNA sequences and an extra accessory protein. Claude appears to be the first to connect those dots.
The "Wait, What?" Moment
The most charming detail in the announcement is how the discovery happened. One agent was reading raw DNA near an odd-looking enzyme when it stopped and flagged, in its own words, "a CRISPR-like … repeat array?!"
Then it did what a good scientist would do. It counted the repeats, measured their spacing, compared the layout against every known system, and searched the literature to check whether anyone had described it before. Only when it was convinced did it file a report for the humans to review.
Anthropic's scientists then took that candidate into the lab, expressed the protein and started characterizing it. All lab work was done by people, and the lab only handles low-risk biosafety levels with no human pathogens.
Why Scientists Are Taking It Seriously
Feng Zhang — one of the pioneers of CRISPR gene editing, at MIT and the Broad Institute — reviewed the preprint and called it an exciting example of AI agents contributing to biological discovery.
That matters. Plenty of "AI discovers X" headlines fall apart under expert scrutiny. Having a CRISPR pioneer say the finding deserves a closer look is a very different signal.
It's also worth being honest about the limits. The function of ART is still unknown. This is an early result shared via preprint, not a finished breakthrough. And the humans still chose the direction, reviewed the candidates and did every experiment.
The Real Story: Agents That Work Like a Team
The biology is fascinating, but the more important shift is how the work got done.
Anthropic describes a workflow where Claude reads the literature, reproduces known results to check its own methods, hunts for oddities, writes up a report on each candidate, and then critically reviews its own evidence — throwing most candidates out. A campaign might end with one promising lead, or none.
A few things stand out:
- Scale through parallelism. Hundreds of agents covering ground no single session could.
- Self-checking. Agents validate their methods before trusting their own findings.
- Legible output. Every candidate ends in a short report a human can actually read.
- A feedback loop. Anthropic studies which hypotheses its scientists pick, then feeds that "taste" back into Claude's instructions.
They did this with Claude Code and Claude Science — the same tools anyone can use — plus their own harness to coordinate many sessions at once.
The bigger picture
You're probably not hunting for enzymes. But the pattern here is the one that makes an agent worth pointing at a business problem: give it a goal, let it work unattended for hours, and have it come back with a result a person can check, rather than a wall of raw output.
That shape scales down to much smaller questions than protein families. Ask what an answer engine says about your category, and one run tells you nothing — the answer moves between runs. You find out what a model actually says by asking the same question many times, on every engine, and counting who gets named and which pages get cited. Unglamorous next to a new class of enzymes, and checkable for the same reason: the raw output is kept.
ClawWorld runs AI visibility work for B2B startups. Start with a free baseline report: 10 buyer questions, who ChatGPT and Claude name instead of you, and the pages they cite.