About 950 agents searched for 21 hours. People kept the lab.

On September 23, Anthropic said a group of Claude agents had searched a large DNA database, flagged a bacteriophage enzyme system, and handed the finding to human scientists. The company calls the system array-associated reverse transcriptases, or ART. It also says the system’s function is still unknown. The search and the lab were different jobs. For the standing version of that split, see The AI-first company, for owners.

What Anthropic reported

Anthropic says it formed a life sciences group in spring 2026 and built a lab in the Bay Area. The group looks through DNA for protein families that are still uncharacterized. Agents propose what a family might do. People test the proposals that survive review.

This search started from one prompt: look through DNA sequences for new reverse transcriptases, the enzymes that copy RNA into DNA. Anthropic says its scientists wrote that prompt. The agents then combed the database, compared families, and chose what to flag. One agent filed a report for a person to review. People did the lab work.

The counts are Anthropic’s. About 950 agents spent 21 hours and a company-reported 210 million tokens. They gathered more than 200,000 reverse transcriptases, marked 3,500 candidate systems, and wrote reports on the 20 they found most compelling. Anthropic says an expert scientist can spend weeks or months on that kind of pass.

One agent, reading the DNA beside an unusual reverse transcriptase, noticed a repeating pattern. Anthropic says that enzyme was already in the literature, in a jumbo phage. The new observation was what sits beside it: an array of non-coding DNA that looks like CRISPR, and an accessory protein whose function is also unknown.

Anthropic says human scientists do the lab work. It says the first experiments show that the array is expressed as a set of distinct short RNAs, that this suggests something analogous may be at work, and that the job of ART is still open. Anthropic links a pre-print.

How a pass like this is staffed

Anthropic says the agent reads the literature on a family and tries to reproduce results that are already known, from public data, as a check on its method. It then looks for relatives, or for neighbors in the genome, that fit no described system. Each one gets a short report a person can read: a proposed function, and the evidence for it. A later pass throws most of them out. Anthropic says a survey can end with one candidate worth testing, or with none.

When a candidate survives review, Anthropic says people test it in the lab. They express the protein in standard laboratory strains and characterize it, and Claude helps interpret the data. The search ran in Claude Science and Claude Code, tools Anthropic says any scientist can open, sometimes with a harness the lab built to run many sessions at once.

Anthropic says one campaign can yield hundreds or thousands of reports. The scientists study which proposals they will test, and they write that standard back into the instructions.

Four jobs stayed with the people in that account:

1. The question. One prompt, pointed at a pile too large to read by hand this week. Theirs named a family of enzymes, a kind of database, and a kind of oddity.
2. The known case. The agent had to reproduce results that were already settled before it went looking for something new. Match those cases before the agent sees the rest of the pile.
3. The shortlist. The agents cut 3,500 candidates to 20 written reports. Anthropic says the agents’ later pass usually eliminates most proposals. A person reads that shortlist.
4. The step that is hard to undo. Running the experiment stayed with people who can stand behind the result. The core gene was already known, and the function is still open.

Run it on a pile you already have

Pick one pile you will not finish reading yourself this week. Old proposals, support threads, call notes, or contracts you meant to revisit. Write these five lines before any agent touches the full pile. If a line stays blank, wait.

1. Question. One sentence a stranger could apply to a single file, such as whether a proposal names a problem you still have not solved. A request to find anything useful fails that test.
2. Known case. One file you have already judged, and the answer you gave it. The agent reads that file first. If it misses your answer, change the question before it sees the rest.
3. Shortlist. A count you can read in one sitting. Anthropic’s agents stopped at 20 full reports. Pick your number before they start, and hold them to it.
4. Kill rule. What sends an item back: no line taken from the source file, a maybe with no evidence, or a topic outside the sentence in line 1.
5. Your step. The action that is hard to undo. A price, a send, a filing, a promise, or a change to a live system. It waits until you have read the shortlist.

Each survivor comes back as a short note: the answer, a pointer to the file, and what would make the note wrong. You read that stack.

What to do this week

1. Choose the pile and write the five lines. Keep them with the files.
2. Run the known case first. Stop if the agent cannot match the judgment you already made.
3. Read the shortlist yourself. Keep the notes that point at a source. Discard the rest.
4. On anything you keep, write one sentence on what the file shows and what it does not show yet.

Writing that question, and deciding which notes deserve an hour, is the work we do with owners in a consult block. The agents can read the pile. You decide which notes get the next hour.