Multi-Agent Orchestration with NVIDIA NeMo Agent Toolkit

Why it matters

Agent-as-tool is the simplest way to give specialists their own tools and prompts. The notebook shows how much configuration that buys, and where control actually sits.

What to practise

Register retail-sales tools, then write a config whose top-level react_agent has no tools of its own: its tool_names are three specialist agents — a tool_calling_agent for data analysis, a visualization agent and a RAG agent — each with its own max_iterations and description. Run several questions with nat run and watch which specialist the orchestrator picks. Then replace the visualization agent with a custom LangGraph agent that asks a human before summarizing a chart.

How it fits

This is the agent-as-tool topology from 1.5.1, with the orchestrator owning control. Compare it against the explicit handoffs and failure outcomes in 1.5.2, and ask the question 1.5.3 asks: would one agent with all the tools score as well at lower latency and token cost?