Adding Tools to NVIDIA NeMo Agent Toolkit

Why it matters

An agent picks a tool from its name, description and parameters alone. Writing tools by hand shows how much of that contract you control and where it is left loose.

What to practise

Scaffold a workflow with nat workflow create, then write three tools over a retail sales CSV. Each has a FunctionBaseConfig subclass, a @register_function(config_type=...) builder and an inner function returned through FunctionInfo.from_fn(), its docstring serving as the description. Import them in register.py, list them under functions and the react_agent’s tool_names, and use nat run to check that each question reaches the expected tool. Then add a LlamaIndex retrieval tool with an embedder. Note what the docstrings leave unsaid, such as that the sales tools match product names only in lowercase.

How it fits

Judge each tool’s name, description and return type against 2.3.1. The tools read a local file without checks; 2.3.2 is the validation and authorization boundary a tool that changes data would need.