Skip to content
Neural Daily

ToolsNews

NVIDIA Agent Toolkit grounds agents in company data

At GTC Berlin NVIDIA expanded its open source Agent Toolkit. The new Agent Knowledge Foundation is meant to let agents work from a company's own data without moving it.

By: Mikola Máté Patrik · · 2 min read

NVIDIA Agent Toolkit illustration: a prompt box feeding a central agent panel linked to data, model and tool icons over an office background
Image: NVIDIA · Press kit · Official blog social image (og:image) · Source

What changed

On October 21 at GTC Berlin, NVIDIA added new pieces to the open source NVIDIA Agent Toolkit (NVIDIA blog). The toolkit bundles models, runtime, data layer and tools into one agent stack. We already covered the Kumo and Nemotron Retriever models in our GTC Berlin open models story, so this one focuses on the agent side.

  • Agent Knowledge Foundation: a new layer that turns company data into knowledge agents can use (NVIDIA Developer). NVIDIA says it works without moving data, rebuilding existing workflows or changing governance controls.
  • Auto-Ontology: a new open library that maps raw data to business concepts such as customers, products and orders, and maps how they relate (GitHub).
  • CUDA-X libraries underneath: cuDF, cuGraph, cuML, cuOpt and cuVS.
  • Kumo Structured models in four variants: Tabular, Relational, Time Series and Forecast (Kumo). Agents can use them to get predictions straight from tabular data.
  • OpenShell, a secure agent runtime that controls what agents can access and do (NVIDIA).
  • A new NemoClaw blueprint for agent optimization, a step by step starting point for improving agents over time (build.nvidia.com).
  • New tools and skills: NeMo Relay and NeMo Compass. The announcement gives no details on these.

Who is using it

NVIDIA's own supply chain team uses Agent Knowledge Foundation through Palantir Foundry. Databricks pairs Kumo Anomaly with Genie Ontology, and Snowflake Notebooks can run Kumo Engine and Kumo Relational. NVIDIA says SAP's fine-tuned Nemotron Retriever reranker ran nearly 3x faster, and ServiceNow cut latency by nearly a quarter while improving reranking quality. These are vendor figures without independent benchmarks.

Why it matters

The weak spot of enterprise agents is rarely the model. It is that they do not know the company's data and vocabulary. Auto-Ontology targets exactly that layer: if it reliably detects entities like customers, products and orders and how they connect, it can save a lot of hand-written RAG and schema description work. It is open source, so try it on one of your own databases before committing to the full NVIDIA stack.

    Produced with the help of AI, reviewed by an editor. How we make our articles

    Reviewed and edited by the Neural Daily desk. About

    Related articles