Artificial Intelligence

NVIDIA Expands Agent Toolkit to Ground Enterprise AI

Leading enterprises are adopting the expanded open-source NVIDIA Agent Toolkit and Agent Knowledge Foundation to customize, control, and ground AI agents in proprietary company data.

Radwan AbdulhadiEditor and technology writer

Published: October 21, 2026 · Updated: October 21, 2026 · 3 min read

NVIDIA Expands Agent Toolkit to Ground Enterprise AI — Image: NVIDIA Blog

Expanding the Open Agent Stack

NVIDIA has announced an expansion of the open-source NVIDIA Agent Toolkit, providing enterprises with new tools to customize, control, monitor, and govern AI agents. Because agents are only as useful as what they know and only as trustworthy as their operational boundaries, builders require an open agent stack that unites models, harnesses, runtimes, data, and tools. This comprehensive framework enables developers to build agents that accomplish real work securely from development to deployment.

To help organizations turn raw data into company intelligence, NVIDIA introduced the NVIDIA Agent Knowledge Foundation. This new foundation allows agents to find relevant information, understand its business context, make predictions, and guide human decision-making. Crucially, companies can deploy the foundation without moving their data, rebuilding existing workflows, or altering their established governance controls.

CUDA-X Libraries and Open Models

The Agent Knowledge Foundation incorporates several NVIDIA CUDA-X libraries, including cuDF, cuGraph, cuML, cuOpt, and cuVS. It also introduces the Auto-Ontology library, which translates raw corporate data into familiar business concepts such as customers, products, and orders while mapping the relationships between them.

Additionally, the foundation features open models including the NVIDIA Nemotron Retriever and the new Kumo family. The Kumo structured models comprise Kumo Tabular, Kumo Relational, Kumo Time Series, and Kumo Forecast, all designed to assist agents in retrieving relevant company data to answer questions and generate accurate predictions.

Secure Runtimes and Blueprint Designs

To maintain strict operational security and control, the toolkit relies on OpenShell to provide a secure runtime boundary. This environment gives enterprises precise oversight over what agent tools can access and execute.

Furthermore, NVIDIA NemoClaw blueprints deliver reference designs for building secure and specialized agents. The latest NemoClaw blueprint focuses on agent optimization, offering a step-by-step starting point that enhances agent intelligence and performance over time.

Early Enterprise Adoption

Global enterprises are already integrating the NVIDIA Agent Knowledge Foundation into their production environments. NVIDIA's own supply chain organization is partnering with Palantir to deploy specialized agents, generate actionable intelligence from proprietary data, and optimize decision-making processes.

Other major technology providers are incorporating the new capabilities into their platforms. Databricks is pairing NVIDIA Kumo Anomaly with Databricks Genie Ontology to support context-aware agents, while Snowflake allows customers to leverage the NVIDIA Kumo Engine and Kumo Relational open models within Snowflake Notebook environments.

Search and data platforms are also reporting substantial performance improvements. Amazon OpenSearch Service enhanced its search relevance with fewer steps and virtually no added latency, SAP's fine-tuned reranker ran nearly three times faster, and ServiceNow improved its reranking quality while reducing latency by nearly a quarter. Meanwhile, Zoom is utilizing Nemotron Retriever models to optimize document understanding and search relevance across its AI services.