ReactiveByDesign · September 24, 2026
On September 23, 2026, Anthropic published a preprint titled Autonomous AI agents discover reverse transcriptases with tandem repeat arrays (Yoon et al., 2026). The paper describes a multi-agent campaign in which Claude agents autonomously searched 1.9 billion protein clusters and identified a previously uncharacterized enzyme system in bacteriophage DNA. The authors included verbatim excerpts from the agent's reasoning transcripts.
On page 5 of the preprint, the following passage appears verbatim from the agent's reasoning trace during the discovery session (task t0062):
"Actually wait — reconsider. Repeat+spacer arrays immediately 5' of an RT in phages… This is EXACTLY the architecture of 'retron-like tandem arrays' in the recently described 'DRT9'? I recall 2024-25 papers on phage/bacterial RTs that polymerize from repeat arrays, e.g., 'ApeA'? Hmm, and there's the Zhang lab's 2024 'programmable RNA-guided…' no. The honest path: characterize quantitatively, then run the novelty kill-test literature search with precise descriptors […]."
— Worker agent, task t0062 · Yoon et al., 2026, p. 5
In this passage, the agent cites "ApeA" as an example of phage/bacterial reverse transcriptase systems that polymerize from repeat arrays. ApeA is a published antiphage defense system. Its documented mechanism is RNA cleavage via a HEPN RNase domain. The published literature does not attribute reverse transcriptase function or repeat array polymerization to ApeA.
The agent's own next sentence acknowledges that the literature check had not yet been performed. A full search of the 40-page preprint confirms ApeA does not appear again anywhere in the document — not in the results, discussion, methods, supplementary figures, or 103-entry reference list. No correction appears in the same document.
UCA — Unverifiable Claim Assertion
A category of our own creation. A specific claim introduced into the record before the verification the agent itself identified as necessary, with no correction appearing in the same document. Confidence level — whether tentative or assertive — is not a determining factor.
We find this instance interesting because it appears in a public, author-released transcript, the agent itself signals that verification was pending, and no correction appears anywhere else in the same document.
This note makes no claim about the significance of the ART discovery, the quality of the research, or the reliability of AI agents in scientific work. It records one passage we found worth noting, nothing more.
[1] Yoon, P.H. et al. (2026). Autonomous AI agents discover reverse transcriptases with tandem repeat arrays. Anthropic preprint. https://www-cdn.anthropic.com/22573675ada52a8ca8a97a1a4b4326b2f208a071.pdf
[2] Anthropic. (2026, Sept. 23). Claude discovers a novel enzyme system with CRISPR-like repeats. https://www.anthropic.com/news/claude-discovers-novel-enzyme-system
[3] Gao, L. et al. (2020). Diverse enzymatic activities mediate antiviral immunity in prokaryotes. Science, 369(6507), 1077–1084. https://doi.org/10.1126/science.aba0372
[4] Drobysheva, A. et al. (2026). ApeA cleaves genomic RNA to defend against RNA phage infection. bioRxiv. https://doi.org/10.64898/2026.03.20.713152
[5] Juozapaitis, J. et al. (2026). Deoxydinucleotides activate the bacterial anti-phage defense system ApeA. bioRxiv. https://doi.org/10.64898/2026.01.26.701840
This note records an observation of interest only. It does not constitute a formal audit, accusation, or claim against any party.
ReactiveByDesign · Finding of Interest · September 24, 2026 · UCA — Unverifiable Claim Assertion
