NewsAnthropic1 min read
Claude Agents Discover a Previously Unknown Enzyme System
About 950 Claude agents combed a DNA database and found an enzyme system with CRISPR-like repeats. What was found and what is still unknown.
Key points
- About 950 Claude agents worked for 21 hours and used 210 million tokens.
- They found an enzyme family in viruses that infect bacteria, with CRISPR-like repeats.
- Humans only gave the initial prompt and ran the lab experiments.
- It is a preprint without peer review, and the system's function is still unknown.
On 23 September Anthropic published a result that shows what a multi-agent system can do with an open scientific problem. Its agents found, in a huge DNA database, an enzyme system nobody had described before.
What the agents did
The team asked Claude to look for interesting examples of reverse transcriptases, enzymes that copy RNA into DNA. From there, about 950 agents worked for 21 hours and used 210 million tokens: they went through the database, studied the different enzyme families and decided on their own which candidates deserved attention. According to Anthropic, human involvement was limited to the initial prompt and the lab work.
What they found
The finding is called ART (array-associated reverse transcriptases). It appears in bacteriophages, the viruses that infect bacteria, and combines three parts: the enzyme, a partner gene and a run of repeated DNA sequences reminiscent of CRISPR.
That combination matters because until now it had only been seen in a handful of systems, all programmable and able to cut, copy and paste DNA. That is why the result has drawn interest: Feng Zhang, one of the CRISPR pioneers, calls it “genuinely intriguing”.
What is still unknown
- What it does. Anthropic acknowledges it does not yet know the system’s main function and is still investigating.
- Whether it holds up. It is a preprint with a technical report, not a peer-reviewed paper.
- Whether it becomes a tool. Resembling gene-editing systems does not mean it can be used as one.
Why it matters
Beyond biology, the story shows the pattern we cover in agent design patterns: many agents working in parallel on a large problem, with people setting the goal and checking the result in the real world. It is expensive in tokens, but in research it can pay off.
