Context
Pezzuti and colleagues publish the Agentic Search Queryset (ASQ), approximately one million traces from open-source agentic RAG systems across three corpora, five agent variants and six retrieval settings. In the controlled experiments, an initial query could produce up to 186 retrieval calls, and query-reformulation patterns differed from those documented for human search.
What changed
The October 21, 2026 PLOS ONE paper releases a peer-reviewed, reproducible dataset and analysis of multi-step agent retrieval, with publicly available ASQ traces and accompanying code.
Why it matters
AI agents can issue numerous evolving retrieval queries for a single user information need. Measuring only the initial query can miss intermediary source exposure and retrieval workload. This evidence supports treating query fan-out, retrieved source exposure, and final citations as separate stages.
Visibility OS interpretation
Visibility OS interpretation: Agentic search discovery cannot be reliably modeled as a single human-style query. The observed maximum of 186 calls is benchmark-specific; this study does not establish commercial AI engine query fan-out, citation policies, or publisher ranking.
What remains uncertain
How closely do these open-source agent traces resemble commercial AI search workflows? How do retrieval fan-out and reranking affect which sources are ultimately cited or used? What fraction of intermediate retrieved pages influences final answers?