NVIDIA Deep Researcher Agent Blueprint

Why it matters

A blueprint shows how NVIDIA itself splits a multi-agent application into deployable services. Reading one is faster than inventing the boundaries yourself.

What to practise

Run the blueprint, built on the NeMo Agent Toolkit and LangChain Deep Agents, three ways: nat run with configs/config_cli_default.yml, start_e2e.sh for the backend API and web UI, and docker compose in deploy/compose, choosing a backend through BACKEND_CONFIG. Watch the orchestration node classify intent and pick shallow or deep research, and compare the YAML configs that swap models and data sources without code changes. Note the Helm assets beside the Compose ones, the async research jobs, and the nat eval run that generates reports for Deep Research Bench.

How it fits

The split between orchestration, research agents, model endpoints and the frontend is 4.1.1’s service boundary. The Compose and Helm assets are 4.1.2’s choice, and the async jobs with durable outputs are the pattern 4.1.3 asks you to make recoverable.