AI has already changed how scientists predict protein structures and design new molecules. The next leap may be discovery itself, with agents that read raw biological data, spot what looks strange, and chase it down on their own. A new study offers an early glimpse of that future, and of what it could mean for the hunt for the next generation of therapeutic tools.
Researchers at Anthropic reported that a team of autonomous large language model (LLM) agents surveyed reverse transcriptase (RT) genes across 1.9 billion metagenomic protein clusters and surfaced a new family of the enzymes, along with the repeat arrays that define it.
RTs copy RNA into DNA, an activity that underpins cDNA synthesis, prime editing, and retron-based genome engineering. Most recently described bacterial RT systems came from genome mining, where experts comb sequence databases and flag anything unusual. That manual curation step struggles to keep pace with metagenomic databases that now double roughly every four years, according to the authors.
How the agents worked
The team built a harness in which worker agents planned and ran analysis tasks while supervisor agents reviewed the results and opened follow-up tasks. The campaign ran for 21.5 hours without human intervention, spanning 949 agent sessions and 119 tasks, 98 of which the agents opened themselves.
The agents recovered about 200,000 RT clusters, sorted them into nine classes, and scored 3,564 protein families that recurred near RT genes as candidate partners. Of the 17 candidates promoted for deep dives, three held up as previously unreported RT associations. The agents set aside the remaining 14 as annotation artifacts, parts of known systems, or general genomic neighbors.
A repeat spotted in raw sequence
The most notable finding came from a detour. One worker dismissed a candidate partner gene as a spurious association but queued a follow-up on the RT itself. The next worker pulled the upstream DNA of related RTs into its context and recognized a pattern directly from the sequence. "I can see by eye a tandem repeat array," the agent wrote, according to the session transcript published in the paper. After a literature search turned up no match, the agents filed a report for human review.
In follow-up analyses, the team defined a family of 95 RT clusters in jumbo phages and predicted viral genomes, 28 of which carried a detectable repeat array upstream of the RT. The researchers named them array-associated reverse transcriptases (ARTs).
The arrays hold three to 21 copies of a repeat 15 to 49 nucleotides long, separated by spacers of 120 to 220 nucleotides. They resemble CRISPR arrays, though the spacers run several times longer, stay conserved between related phages, and sit in loci that lack cas genes. The RTs also carry an unusual N-terminal extension of about 180 residues and pair with one of three unrelated partner protein types.
Abundant RNAs, unknown function
Reanalyzing published RNA sequencing data from Staphylococcus phage SA1 infecting Staphylococcus lentus, the team found that array-derived RNAs made up as much as eight percent of phage RNAs 15 minutes after infection. When the researchers expressed the SA1 system on plasmids in Escherichia coli, the array produced discrete short RNAs similar to those seen during infection.
The authors proposed that ARTs work like retrons, bacterial defense systems built from an RT, a noncoding RNA, and a partner protein, but with a bank of distinct RNAs in place of a single one. Whether the RT is active, whether it interacts with its partner, and what the system does for the phage all remain open questions awaiting bench work.
Lessons for teams deploying AI agents
The discovery proved hard to reproduce. In ten reruns of the same campaign, agents sampled ART loci but missed the array every time.
In fixed benchmarks, the most capable models described the array in at least 90 percent of attempts when the DNA sat directly in their prompt. With sequences supplied as files alongside analysis tools, detection fell as low as 32 percent, largely because 39 percent of those attempts read fewer than 200 contiguous nucleotides of sequence. Detection climbed as models read more raw DNA.
For R&D groups building agentic workflows, the results suggest that harness design and how directly agents engage with primary data can shape what they find. The authors also pointed to a gap human curation left open. The original genome report for phage MarsHill identified its RT and proposed an upstream noncoding RNA, while the repeats and partner gene went undescribed.
