Problem:
Nemotron 3 Super is trained for agentic tasks across 21 RL environments (TerminalBench, TauBench V2, SWE-Bench) and scores 85.6% on PinchBench. However, the repository has no use-case example demonstrating structured tool calling with Super 3.

The existing agent examples use older models (Nano 9B via Streamlit, Nano 12B VL for RAG) and don't showcase the structured function-calling capability that makes Super 3 competitive with frontier models for agentic workflows.

Proposed Solution:
Add a Jupyter notebook use-case example in use-case-examples/Agentic-Tool-Calling-with-Nemotron-Super/ that demonstrates:

1. Defining tools with JSON schema descriptions
2. Single and multi-turn tool calling via vLLM's OpenAI-compatible API
3. A full agent loop where the model plans, calls multiple tools in sequence, and synthesizes results
4. Using Super 3's reasoning modes (reasoning-off, regular, low-effort) with tool calling
5. Best practices for system prompts and handling edge cases

The example will use simulated tools (no external API keys beyond NVIDIA) so it's self-contained and reproducible, following the pattern established by the voice-rag-agent example.

Why now:
With NemoClaw launching at GTC and Greptile's independent evaluation confirming Super 3's tool-calling strength, there's natural community interest in seeing official examples.

I'm willing to implement this. Happy to adjust based on feedback.