AI Agents & Scientific Research
Anthropic preprint reports agents found a new phage RT family
A large computational campaign surfaced an RNA-associated reverse-transcriptase family, but ten reruns missed the defining array and its function remains unknown.
What happened
An Anthropic-authored preprint reports that a Mythos 5 agent campaign surveyed reverse-transcriptase loci across 1.9 billion protein clusters and identified a phage family beside tandem repeat arrays and partner genes, which the authors call ART. The campaign used 119 tasks, 949 agent sessions, 215.6 million tokens and 21.5 hours of wall-clock time. Human scientists reviewed the result and performed RNA-sequencing experiments that found discrete short RNAs from the array. Ten reruns of the same campaign did not rediscover the defining array. The preprint has not been independently peer reviewed. It does not show that the reverse transcriptase is active, that the short RNAs are its substrates, that the partner interacts with it or what the system does for the phage.
Why it matters for AI
The study is a concrete example of agents searching primary scientific data and surfacing an anomaly that experts can test. Its failed rediscovery runs also show why one promising result cannot be treated as a repeatable autonomous-discovery system.
AIRunsThis View
For agent-assisted research, preserve the exact data, tools and run traces, repeat the search, compare failed runs and require domain-expert review plus physical validation before raising confidence. Report what the system noticed separately from what experiments established.
What this could mean for us
A useful design reference for bounded research workflows. AIRunsThis has not reproduced the result, adopted the system or authorized biological or medical work.
Source and method
This item is based on a primary source. AIRunsThis separates the source’s factual announcement from its own interpretation. Vendor performance claims have not been independently reproduced unless stated.