NVIDIA expands Agent Toolkit with enterprise data

On October 21, 2026, at 10:10 UTC, NVIDIA announced an expansion of the Agent Toolkit with a corporate knowledge base, open models and security features for enterprise agents.

What it is and what it isn't: an open stack to build agents connected to company data, not a new single model. (NVIDIA)

In this article you will see:

• what the Agent Knowledge Foundation does
• which models and libraries are in the stack
• how runtime and blueprints handle governance

NVIDIA Agent Knowledge Foundation connects agents to data

NVIDIA says that the Agent Knowledge Foundation lets agents find corporate information, understand its context and support forecasts and decisions, without moving data or replacing existing flows and controls.

The foundation combines five CUDA-X libraries, cuDF, cuGraph, cuML, cuOpt and cuVS, with Auto-Ontology. The library transforms raw data into business concepts, such as customers, products and orders, and maps the relationships between them. (NVIDIA)

Credit: NVIDIA.

Nemotron Retriever and Kumo bring retrieval and forecasts

The expansion includes the open model Nemotron Retriever and the Kumo Structured family. NVIDIA lists four Kumo models: Tabular, Relational, Time Series and Forecast.

The Retriever helps retrieve relevant information for agents. The Kumo models work with structured data to support forecasts, recommendations and risk assessment, according to NVIDIA.

OpenShell and NemoClaw add execution controls

OpenShell provides a secure execution boundary to control what each agent can access and do. The NemoClaw blueprints offer reference projects to build specialized agents with defined controls.

The stack also includes tools and skills such as NeMo Relay and NeMo Compass. To see another benchmark published by NVIDIA, read Vera and CoreWeave's test on agent tasks. The agent supervision proposal also appears in Microsoft's Guardian Agents.