NVIDIA Expands Agent Toolkit With Data, Runtime and Optimization Tools

NVIDIA expanded its open Agent Toolkit on Oct. 21 with data, retrieval and runtime software for building enterprise agents around internal information and defined access controls.

The new NVIDIA Agent Knowledge Foundation is intended to let companies use their data without moving it, rebuilding existing workflows or changing governance controls. It combines CUDA-X libraries cuDF, cuGraph, cuML, cuOpt and cuVS with a new Auto-Ontology library. Auto-Ontology organizes raw data into business concepts such as customers, products and orders, then maps how those concepts relate.

The foundation also incorporates Nemotron Retriever and four open Kumo Structured models: Kumo Tabular, Kumo Relational, Kumo Time Series and Kumo Forecast. They are intended to help agents find relevant company information and make predictions. The toolkit’s listed tools and skills also include NeMo Relay and NeMo Compass.

OpenShell provides a runtime boundary for controlling what agents can access and do. NemoClaw supplies reference blueprints for specialized agents, with its newest blueprint focused on improving agent performance over time. The runtime work extends NVIDIA’s earlier OpenShell security tools.

NVIDIA’s supply chain organization uses the foundation through Palantir Foundry, Ontology and AIP to deploy specialized agents and draw insights from proprietary data. Databricks pairs NVIDIA Kumo Anomaly with Databricks Genie Ontology. Snowflake customers can use the Kumo Engine and Kumo Relational open models in Snowflake Notebook environments.

The release also includes performance examples. Amazon OpenSearch Service improved search relevance with fewer steps and virtually no added latency using an embedding model. SAP’s fine-tuned reranker ran nearly three times faster, while ServiceNow improved reranking quality and cut latency by nearly a quarter.