GxP-Agent uses graph topology for 100% reliable clinical trial data

Jaime Yan introduces GxP-Agent, a multi-agent system designed to tackle the persistent challenge of clinical trial programming under CDISC standards. Traditional LLM-based approaches fail catastrophically on this task—none of five frontier models produced a valid subject-level analysis dataset in 11 single-shot attempts. GxP-Agent addresses this by encoding regulatory process ordering as a directed acyclic graph (DAG), breaking down monolithic dataset generation into 15 domain-specific nodes executed by specialized worker agents with pharmaverse skill context and validation gates.

On the new CDISC-Bench benchmark derived from the FDA’s CDISCPilot01 (254 subjects, 49 ground-truth ADSL variables), GxP-Agent powered by Claude Sonnet 4.6 achieved 100% structural match across three independent runs—matching all 49 variables and 254 records perfectly. This compares to just 59.2% for the best retrieval-augmented baseline and 0% for all single-agent and flat multi-agent approaches. The graph topology even enables weaker models like GPT-4.1 to achieve 59.2% structural match under the same DAG, whereas it scored 0% under all other architectures. The system generalizes to other datasets, such as ADAE (adverse events), achieving 100% structural match on the first attempt with a 9-node branching DAG and 1,191 records.