Claude agents found a new enzyme system in 21 hours. Humans decided it was real

Anthropic said on Wednesday that Claude agents, left to run without human intervention, found a family of enzymes nobody had described before. The company calls it ART, for array-associated reverse transcriptases: an enzyme found in large bacterial viruses, a partner gene beside it, and a long run of evenly spaced DNA repeats that looks like the arrays behind CRISPR gene editing. The preprint puts the search at 949 agent sessions, 21.5 hours and 215.6 million tokens across 1.9 billion protein clusters.

The headline reading is that AI now does science. The preprint says something narrower and more useful for anyone whose work is research.

The agents checked their own leads. People checked the agents

The campaign ended with 17 candidate gene families. Only three were confirmed as new. The other 14 were rejected or set aside as annotation artefacts, parts of systems already described, or genes that simply happened to sit nearby. Agents did that sorting too: supervisor agents reviewed every task, 49 of the 119 tasks were sent back at least once, and an editor agent reviewed each report before it reached a human.

The find itself was hard to repeat. Anthropic ran the same campaign ten more times with the same brief. Most runs touched the ART lineage. None read the DNA next to the enzyme, and none found the array. In a controlled test, the most capable models described the array correctly in at least 90% of attempts when handed the DNA directly, and in as few as 32% when they had to fetch it with tools.

Then the lab. Anthropic says all of the lab work is done by human scientists, at a Bay Area facility limited to lower-risk biosafety levels. They expressed the proteins, sequenced the RNA and characterised the structures. What ART does is still unknown. Dario Amodei told TechCrunch that Stanford researchers had already found a system that is in some ways similar, so the field has yet to decide how new this is.

Feng Zhang, the CRISPR researcher at MIT and the Broad Institute, called the finding "genuinely intriguing" and said it "merits further investigation". That is how a working scientist talks about a lead, not a result.

What moves in a research job

The volume layer has moved. Reading through a database no team could cover, writing up candidates, opening follow-up tasks: 949 sessions did that overnight. If your week is mostly that layer, it is the part of the job a system now performs.

Part of the checking layer has moved as well. Agents reviewing agents rejected 14 of 17 leads before a person read a word. What stayed human is the rest of the judgement: writing the brief, deciding which report deserves bench time, noticing that ten reruns never read the DNA that mattered, and designing the experiment that settles it. Anthropic says the scientists' feedback goes back into the instructions to teach Claude "our own scientific taste". Taste is the scarce input here, and it is still supplied by people.

AI does the routine work. You do the thinking. That is the split this campaign shows. For the method applied to your own role, see Will AI take your job?.

Your Next Move

• Audit your search work. List the tasks in your week that are reading, screening or summarising at volume. Those are the ones an agent campaign now covers.
• Move up to the proof. Screening is now partly done by agents checking agents. The value left sits in choosing what to test and designing the experiment that confirms it. Get visibly good at deciding what is real.
• Treat one agent run as one sample. Ten reruns missed what the first one found. Run agent work more than once and check what each run actually read.