NVIDIA Expands Agent Toolkit With New Knowledge Foundation and Open Models
NVIDIA adds new tools and open models to its Agent Toolkit, enabling enterprises to build agents grounded in proprietary data and governed by strict security boundaries.
Key points
- The NVIDIA Agent Knowledge Foundation integrates CUDA-X libraries and new Auto-Ontology capabilities to translate raw data into business concepts.
- New open models, including the Kumo family and Nemotron Retriever, help agents find relevant information and make predictions without moving data.
- Partners like Palantir, Databricks, Snowflake, and AWS are already integrating these components into their respective platforms for enterprise use.
NVIDIA has expanded its open source Agent Toolkit with a new Agent Knowledge Foundation and additional tools designed to help enterprises build agents that rely on proprietary company data. The update aims to provide builders with a complete stack for developing agents that are both useful and trustworthy from development through deployment.
The new foundation allows companies to turn raw data into actionable intelligence for agents without moving that data or altering existing governance controls. According to NVIDIA’s blog post, this approach helps agents find relevant information, understand its business context, and assist in decision-making while staying within defined operational boundaries.
Core Components of the Update
The Agent Knowledge Foundation features NVIDIA CUDA-X libraries, including cuDF, cuGraph, cuML, cuOpt, and cuVS, alongside a new Auto-Ontology library. This library translates raw company data into familiar business concepts such as customers, products, and orders, mapping how they relate to one another for agent consumption.
The toolkit also introduces new open models to support these agents. These include the NVIDIA Nemotron Retriever and the new Kumo family, which comprises Kumo Tabular, Kumo Relational, Kumo Time Series, and Kumo Forecast models. These models help agents locate relevant company information to answer questions and generate predictions.
Security and Runtime Controls
OpenShell provides the secure runtime boundary for these agents, giving enterprises control over what agents can access and do. NVIDIA also offers NemoClaw blueprints, which serve as reference designs for building specialized, secure agents. The newest blueprint focuses on agent optimization, providing a step-by-step starting point for improving agent intelligence over time.
These tools are designed to ensure agents remain trustworthy by enforcing boundaries on their actions and access. By combining models, harness, runtime, data, and tools into an open stack, NVIDIA aims to simplify the process of creating agents that can accomplish real work safely.
Enterprise Adoption and Partners
Several major technology companies are already using the NVIDIA Agent Knowledge Foundation. NVIDIA’s own supply chain organization uses it through Palantir Foundry, Ontology, and AIP to deploy specialized agents, create intelligence from proprietary data, and optimize decision-making.
Databricks pairs NVIDIA Kumo Anomaly with Databricks Genie Ontology to power context-aware agents. Snowflake enables customers to use the NVIDIA Kumo Engine and Kumo Relational open models in their Snowflake Notebook environments to unlock more predictive capabilities.
Other partners leveraging Nemotron Retriever models include Amazon OpenSearch Service, which improved relevance with fewer search steps; SAP, which saw its fine-tuned reranker run nearly three times faster; and ServiceNow, which improved reranking quality while cutting latency by nearly a quarter. Zoom is using these models to improve document understanding and search relevance for Zoom AI Services.
Additional partners building on the Agent Knowledge Foundation include Amazon Web Services, ClickHouse, Cloudera, Cohesity, Dassault Systèmes, Dell, Elastic, Exa, IBM, NetApp, Oracle, Perplexity, Starburst, and VAST Data. As Threat Times reported in October 2026, NVIDIA has been expanding its open model portfolio with releases like Nemotron 3.5 Super VL and Kumo to support sovereign AI applications.
What should you do?
Security and technology professionals should evaluate how their organizations plan to integrate AI agents with proprietary data. Review current data governance policies to ensure they align with the boundaries these new tools enforce. Monitor vendor updates for the Agent Toolkit to understand how runtime controls like OpenShell apply to your specific use cases.
What to do and how to stay safe: NVIDIA
- Assess current data governance frameworks to determine how they align with AI agent boundaries and access controls.
- Review the NVIDIA Agent Toolkit documentation to understand the capabilities of the new Auto-Ontology library and CUDA-X libraries.
- Monitor vendor announcements for integration updates from partners like Palantir, Databricks, and Snowflake.
- Evaluate the security implications of using open models like Nemotron Retriever and Kumo within your organization's runtime environment.
Step-by-step guide: Mobile Device Security: How to Protect Data Beyond Passwords
General security guidance from the Threat Times newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
What is the NVIDIA Agent Knowledge Foundation?
It is a new component of the NVIDIA Agent Toolkit that helps enterprises turn raw data into company intelligence for agents using CUDA-X libraries and new open models.
Which companies are using the new NVIDIA Agent Toolkit features?
Partners including Palantir, Databricks, Snowflake, Amazon Web Services, SAP, ServiceNow, and Zoom are integrating these tools into their platforms.
Does the Agent Knowledge Foundation require moving data?
No, companies can use the Agent Knowledge Foundation without moving data, rebuilding existing workflows, or changing governance controls.
