NVIDIA wants to give AI agents access to company data without moving it
NVIDIA is expanding Agent Toolkit with Agent Knowledge Foundation, meant to make company data usable by AI agents. SAP, Snowflake and Databricks are on board, but pricing and availability are not disclosed.


NVIDIA is expanding its Agent Toolkit with Agent Knowledge Foundation, a new set of libraries and open models designed to make a company's own data usable by AI agents. According to NVIDIA, businesses can use it without moving data, rebuilding workflows or changing existing governance controls. Pricing and availability for the new components are not disclosed in the announcement.
NVIDIA published the news in a blog post on Wednesday 21 October 2026 at 12:10 Norwegian time (10:10 UTC), during GTC Berlin. On the same day the company launched the OpenShell Reference Architecture for agent security, which hogby.ai has covered in a separate story. Agent Toolkit is open source software.
What NVIDIA is launching
Agent Toolkit is NVIDIA's package for building, governing and monitoring AI agents. According to the blog post it now has five parts: the new Agent Knowledge Foundation with CUDA-X libraries, open models such as Nemotron Retriever and the new Kumo family, the OpenShell runtime, NemoClaw blueprints for secure agents, and tools such as NeMo Relay and NeMo Compass.
The core of the news is Agent Knowledge Foundation. It is meant to help agents find relevant information in company data, understand what it means for the business and make predictions that support decisions. The package uses NVIDIA's CUDA-X libraries cuDF, cuGraph, cuML, cuOpt and cuVS.
The new Auto-Ontology library, according to NVIDIA, translates raw company data into familiar business concepts such as customers, products and orders, and maps how they relate. It is joined by the Kumo models Kumo Tabular, Kumo Relational, Kumo Time Series and Kumo Forecast, which are meant to help agents answer questions and make forecasts from structured data.
NVIDIA is also launching a new NemoClaw blueprint for agent optimization. It is described as a step-by-step starting point for making agents better at their jobs over time.
Who is using it
NVIDIA says its own supply chain organization uses Agent Knowledge Foundation through Palantir Foundry, Ontology and AIP to deploy specialized agents and make predictions from proprietary data.
According to NVIDIA, Databricks pairs Kumo Anomaly with Databricks Genie Ontology. Snowflake is making Kumo Engine and Kumo Relational available to its customers in the Snowflake Notebook environment.
For the Nemotron Retriever search models, NVIDIA cites concrete customer results. Amazon OpenSearch Service improved relevance with fewer search steps and virtually no added latency. SAP's fine-tuned reranker ran nearly three times faster. ServiceNow improved reranking quality while cutting latency by nearly a quarter. Zoom uses the models in Zoom AI Services. The figures are NVIDIA's own and have not been independently tested.
Among others building on Agent Knowledge Foundation, NVIDIA names AWS, Cloudera, Dell, Elastic, IBM, NetApp, Oracle, Perplexity and VAST Data.
What is not disclosed
The blog post does not state pricing, licence terms for the individual models, or when Auto-Ontology and the new NemoClaw blueprint will be available. Nor does it say which hardware is required, beyond the fact that the package builds on NVIDIA's own CUDA-X libraries. The claim that data does not need to be moved is NVIDIA's description, and the post does not explain how access, traceability and privacy are handled when agents read company data.
What leaders should note
- The contest over AI agents is now about access to a company's own data. Map which systems an agent should read from, who owns the data and how access is logged before you choose a platform.
- Several vendors that Norwegian businesses already use, such as SAP, ServiceNow, Snowflake, Databricks, Oracle and AWS, are on NVIDIA's list. Ask them what the partnership means in practice for their products, and whether it changes pricing or hardware requirements.
- The tools are open source but build on NVIDIA's libraries. Consider how tightly you want to tie your data platform and agents to a single vendor, and include that in your risk assessment.
Sources and media
- NVIDIA: NVIDIA Agent Toolkit Expands, Grounding Enterprise Agents in Company Data, blog post published 21.10.2026 at 12:10 Norwegian time
- The source was checked on 21.10.2026.
- Illustration: AI-generated with Grok Bot / hogby.ai