Governing genomic AI in the era of agentic systems

Genomic artificial intelligence (AI) is moving from task-specific models toward systems that can integrate genomic, molecular, and phenotypic information across increasingly complex analytical workflows. There are governance frameworks for the responsible use of genomic AI, largely developed for systems whose data use and analytical behavior are defined in advance. Agentic systems add a new governance challenge because analytical decisions, resource use, and subsequent actions may no longer be fully specified before execution. In this review, we examine how this shift changes the governance of genomic AI and suggest that oversight must extend to the behavior of the analytical system itself. Permissions, data use, provenance, validation, uncertainty, and human authority need to remain enforceable during execution. We introduce AI systems implementation at the Center for Applied Genomics (CAG) of the Children’s Hospital of Philadelphia (CHOP), including current controls, an ongoing experiment using a restricted project-specific environment for agentic analysis, and a phased approach to broader deployment. Agentic genomic AI governance requires a shift from static review of data and models to enforceable control of system behavior during execution.