NVIDIA Expands Agent Toolkit to Ground Enterprise Agents in Proprietary Data

Enterprise agents only deliver value when they can access the right information and operate within strict governance boundaries. NVIDIA's expanded Agent Toolkit addresses both needs by providing an open stack that combines models, data services, runtime controls, and developer blueprints.

The new Agent Knowledge Foundation leverages CUDA‑X libraries such as cuDF, cuGraph, cuML, cuOpt, and cuVS, together with an Auto‑Ontology layer that translates raw corporate data into familiar business concepts — customers, products, orders — and maps their relationships. It also includes the Nemotron Retriever embedding models and the Kumo family (Tabular, Relational, Time Series, Forecast) to enable fast, accurate search and predictive reasoning.

OpenShell supplies a secure runtime sandbox that defines exactly what an agent may read, write, or invoke, while NemoClaw blueprints give teams reference architectures for building specialized, auditable agents. The latest blueprint focuses on continuous agent optimization, allowing performance to improve over time.

Early adopters illustrate the impact: Palantir integrates the foundation with Foundry and AIP for supply‑chain intelligence; Databricks pairs Kumo Anomaly with Genie Ontology for context‑aware agents; Snowflake exposes the Kumo Engine in notebooks; and AWS, ClickHouse, ServiceNow, Zoom, and SAP report measurable gains in search relevance, latency, and throughput.

Why it matters for GPU and AI infrastructure

By keeping data in place and accelerating graph, tabular, and vector workloads on GPUs, the toolkit reduces data movement costs and latency — key factors for any AI‑centric cloud platform. Secure runtimes and governance hooks also align with enterprise compliance requirements, making GPU‑accelerated agent deployments practical at scale.