Anthropic's new life sciences group used Claude agents to find a new enzyme system in phage DNA. They call it array-associated reverse transcriptases (ART). Here is how the work was split between agents and humans.
The numbers
| What | Value |
|---|---|
| Search time | about 21 hours |
| Agents | roughly 950 |
| Tokens used | 210 million |
| Reverse transcriptases gathered | over 200,000 |
| New candidate systems | 3,500 |
| Candidates written up for humans | 20 |
The five steps
- One prompt. Humans asked Claude "to search through a massive database of DNA sequences for interesting new examples of RTs" (reverse transcriptases, enzymes that copy RNA into DNA).
- Gather. The agents pulled in over 200,000 RTs and investigated the distinct RT families.
- Filter. They picked out 3,500 new candidate systems. Claude "critically evaluates the evidence", and most candidates are eliminated at this stage.
- Report. The 20 most compelling candidates became human-readable reports.
- Test in the lab. Human scientists expressed the protein and did biochemical and structural work, with Claude helping to interpret the data.
What ART is
- A reverse transcriptase, a partner gene and an array of evenly spaced DNA repeats.
- The repeat array is "reminiscent of CRISPR".
- First experiments show the array is expressed as a set of distinct short RNAs.
What is still unknown
- Anthropic says: "we don't yet know its function."
- It has not yet been shown that the enzyme is active, or that it works on these RNAs.
- Feng Zhang (MIT and the Broad Institute) called the finding "genuinely intriguing" and said it "merits further investigation".
Who did what
| Humans | Claude agents |
|---|---|
| Wrote the first prompt | Combed the database |
| Did all the lab work | Judged which candidates were interesting |
| Reviewed the reports | Wrote the reports, helped read the lab data |
Tools used: Claude Science and Claude Code, "the same tools available to any scientist."
The takeaway for your own agent work
The pattern is simple: a wide search, a hard filter, a short list of written reports, then human testing. Keep the human checkpoint at the end, where results are expensive to get wrong.