Can AI agents plan radiation therapy without human guidance?

Radiation therapy plans require careful tuning of dozens of parameters to target tumors while avoiding healthy tissue—a task that typically demands human expertise. This work pairs a reinforcement learning agent (which finds good parameter ranges) with an LLM (which reasons through the optimization) so they guide each other: the RL agent prevents the LLM from proposing physically invalid settings, while the LLM learns to prioritize competing objectives logically. Tested on prostate and liver cases with varying complexity, the system matches expert-quality plans in fewer iterations than LLMs working alone, and generalizes across patient anatomy and treatment sites.