A new multi-agent AI system builds regulator-ready clinical trial datasets with perfect accuracy, a task that breaks every leading model working solo.
The researchers behind GxP-Agent tested single-shot code generation across five frontier models and got zero valid datasets in 11 attempts. Their fix maps the regulatory process itself as a directed acyclic graph, splitting the work into 15 domain-specific steps, each handled by a worker agent with validation checks and retry logic. On CDISC-Bench, a benchmark built from the FDA's own CDISCPilot01 pilot submission covering 254 subjects and 49 variables, GxP-Agent running on Claude Sonnet 4.6 hit a 100% structural match across three separate runs. The best retrieval-augmented baseline managed 59.2 percent; every flat single-agent or multi-agent setup scored zero. The same architecture generalized to adverse-event data, matching all 55 variables across 1,191 records on the first attempt.
The real finding isn't the top score, it's what happened to the weaker model. GPT-4.1 went from 0 percent under every other setup to 59.2 percent once it ran inside the same DAG. That suggests the gain comes from encoding regulatory process knowledge into the system's structure, not from throwing a bigger model at the problem. For an industry where a single formatting error can stall a drug submission, that's a more durable fix than waiting for the next model upgrade.
Worth remembering: this is one team's benchmark built from one FDA pilot dataset, not a live submission reviewed by regulators. A 100 percent score on a known, static test set is a promising upper bound, not proof the system survives the messiness of an actual trial.