Clinical trial coding agent passes a structured FDA-style test

A new arXiv paper reports a multi-agent approach to clinical trial programming, a regulatory workflow where one-shot language-model code generation performed poorly.

The task is to turn study protocols into analysis-ready datasets under CDISC standards. In the authors’ tests, five frontier models failed across 11 single-shot attempts to produce a valid subject-level analysis dataset.

GxP-Agent takes a different route. It represents the regulatory process as a directed acyclic graph, splits dataset generation into 15 domain-specific nodes, adds pharmaverse skill context, and uses validation gates with conditional retry. On CDISC-Bench, built from the FDA pilot submission CDISCPilot01, the system using Claude Sonnet 4.6 reached a 100 percent structural match across three independent runs.

That result is promising for regulated data work, but it is still a benchmark result on a defined task. The broader lesson is that domain order, validation, and retries can matter more than asking a powerful model to generate an entire compliant dataset at once.