RadOnc-Agent records 98.79% function selection on 2,600 single-function requests

The framework maps radiotherapy into four phases and exposes 26 callable functions through an LLM controller that enforces schema constraints and maintains longitudinal state. Ablation tests showed removal of state tracking dropped cross-stage completion from 96.50% to 84.00%, while disabling validation raised mismatched dispatches to 95.28%. These metrics were obtained on synthetic scenarios and replayed patient records rather than live clinical systems.

Prior agentic systems such as those described in the Med-PaLM 2 technical report and the 2024 Agent Hospital simulation paper operated within single modalities or short episodic tasks. RadOnc-Agent extends orchestration across decision, planning, delivery and adaptation phases while preserving patient context across calls. The evaluation design isolates controller reliability from backend model correctness, leaving clinical validity untested.

Operational deployment would require integration with hospital PACS, TPS and EMR APIs plus prospective audit of output safety. The paper explicitly states it does not establish clinical utility or prospective benefit. Next steps include controlled pilot studies measuring plan quality and time-to-decision against current multi-specialist workflows.