Schema-Gated Agentic AI for Flexible and Reproducible Scientific Workflows

ABSTRACT

This paper proposes a schema-gated architecture that separates natural language reasoning from strict execution constraints in agentic AI systems. The framework allows large language models to generate flexible reasoning paths while ensuring that tool invocation and workflow execution conform to predefined schemas. The approach improves reproducibility in scientific workflows where deterministic execution and traceability are critical. By enforcing schema validation at the action layer, the system balances creativity in reasoning with reliability in execution.