ToolOrchestra is a research framework from NVIDIA for training large language models to function as autonomous agents capable of orchestrating multiple tools in coordinated workflows. It provides an end-to-end reinforcement learning training pipeline that enables models to learn when and how to invoke external tools in sequence. The project includes both a training framework, a released 8B parameter model on HuggingFace, and a curated dataset called ToolScale for training such agent systems.