[erichocean]

Here you go: https://research.nvidia.com/labs/lpr/ToolOrchestra/

Big models (like Claude Opus 4.5) can (and do) just RL-train this into the main model.