Developed per-sample tool environment routing for the ToolAgentLoop component in the volcengine/verl repository, enabling heterogeneous tool sets to be assigned to each dataset row during multi-turn rollouts. Leveraging Python and YAML for backend development and configuration management, the solution introduced a config-driven mapping that supports dataset-level routing and preloads named environments to optimize runtime efficiency. The design maintained backward compatibility by defaulting to global tool configurations when per-row routing was not specified and enforced server-side configuration paths to enhance security. This work increased flexibility and realism in agent training scenarios while preserving system safety and stability throughout deployment.