CASCADE is an agentic AI framework for trustworthy biomarker discovery in precision oncology

Autonomous AI agents can now run end-to-end scientific analyses, yet no general method ensures those analyses are correct, and one upstream error can propagate across hundreds of downstream results at machine speed. We present CASCADE, an agentic validation framework built on mandatory, quantitatively gated analytical steps. Its central principle, specification-implementation separation, fixes what must be validated and to what standard while leaving how to the agent, so conclusions depend on the protocol rather than the model executing it. Every analysis is preserved as human-readable code with a complete decision log. Derived from a ∼25,000-patient clinical cancer genomics cohort and applied unmodified across three settings (13,424 patients), CASCADE recovered established signal, reported insufficient power instead of a false result where data were inadequate, and reversed a clinically consequential artifact (hazard ratio 0.84, corrected to 1.25). Findings were replicated in two independent multi-institutional cohorts, and three AI systems reproduced them with high fidelity.