Bringing Your Own Agent to NVIDIA NeMo Agent Toolkit

Why it matters

Teams rarely start from a blank toolkit project. This notebook shows how much of an existing agent can move under toolkit configuration, and what each step of the move buys.

What to practise

Start from a LangChain ReAct agent with a Tavily search tool. Scaffold three workflows with nat workflow create: first wrap the agent in a register_function, then expose its parameters as configuration, then replace its model name with an LLMRef and its tool with the built-in tavily function group, keeping framework_wrappers=[LLMFrameworkEnum.LANGCHAIN] so the builder hands back LangChain objects. Finish with a tool_calling_agent configuration that needs no custom code. Run each stage with nat run and note what moved from Python into config.yml.

How it fits

The stages are 7.3.2’s integration paths, from wrapping an agent to rebuilding it from toolkit components. The final llms, function_groups and workflow sections are 7.3.1’s named-component wiring.