NVIDIA expanded its open-source Agent Toolkit on Wednesday, adding software libraries, runtime controls and specialized models designed to connect enterprise artificial intelligence agents directly to company data.
The update introduces the NVIDIA Agent Knowledge Foundation, which incorporates CUDA-X libraries such as cuDF, cuGraph, cuML, cuOpt and cuVS alongside a new Auto-Ontology library. NVIDIA said the Auto-Ontology tool translates raw internal data into structured business concepts, including customers, products and orders, while mapping the relationships between them. The company stated that businesses can implement the foundation without moving stored records, altering governance policies or rebuilding existing workflows.
The toolkit also adds the Kumo family of open structured models—consisting of Kumo Tabular, Kumo Relational, Kumo Time Series and Kumo Forecast—alongside Nemotron Retriever models to assist agents in searching records and generating business predictions. For execution and governance, NVIDIA packaged the OpenShell runtime boundary to restrict what agents can access, and introduced NemoClaw security blueprints, including a starting reference design for continuous agent optimization. The broader stack also includes tools named NeMo Relay and NeMo Compass.
Enterprise software vendors have integrated components of the updated foundation across existing systems. SAP operated a fine-tuned Nemotron Retriever reranker nearly three times faster, while ServiceNow reduced reranking latency by nearly a quarter alongside quality improvements. Amazon OpenSearch Service used the Nemotron embedding model to raise search relevance with fewer search steps and virtually no added latency.
NVIDIA's own supply chain organization deployed the Agent Knowledge Foundation through Palantir Foundry, Ontology and AIP to forecast outcomes and guide operational decisions. Databricks paired NVIDIA Kumo Anomaly with Databricks Genie Ontology, and Snowflake made the Kumo Engine and Kumo Relational models available inside Snowflake Notebook environments. Zoom integrated Nemotron Retriever models to assist with context retrieval and document understanding in Zoom AI Services.
Other enterprise infrastructure providers building on the foundation include Amazon Web Services, ClickHouse, Cloudera, Cohesity, Dassault Systèmes, Dell, Elastic, Exa, IBM, NetApp, Oracle, Perplexity, Starburst and VAST Data. NVIDIA scheduled a technical presentation on building and deploying specialized agents for its GTC conference in Berlin.
