Preprint reports DAG agents improve clinical-trial dataset generation

A new arXiv preprint reports that GxP-Agent, a multi-agent system for clinical-trial programming, achieved a full structural match for an analysis dataset across three CDISC-Bench runs. The system divides dataset generation into 15 domain-specific directed-acyclic-graph nodes with validation gates and conditional retries. The paper reports matching 49 variables and 254 records, compared with 59.2% structural match for its best retrieval-augmented baseline; it also reports a first-attempt full match on an adverse-events dataset.