NVIDIA Expands Its Agent Toolkit to Ground Enterprise Agents in Company Data
NVIDIA expands its Agent Toolkit with knowledge, models, OpenShell, blueprints and skills to ground enterprise AI agents in private data securely.
Key Takeaways
- NVIDIA has expanded its open-source Agent Toolkit to help enterprises build AI agents that use company data within controlled, secure boundaries.
- The new Agent Knowledge Foundation combines CUDA-X libraries, Auto-Ontology, Nemotron Retriever, and Kumo models to ground agents in private enterprise data without moving data or changing existing governance.
- NVIDIA OpenShell and NemoClaw provide secure runtime controls and reusable blueprints, making governance and production safety core parts of agent deployment.
- Early enterprise uses include Palantir, Databricks, Snowflake, and AWS, with reported gains in retrieval, relevance, latency, and predictive capabilities.
Meta description:
NVIDIA Agent Toolkit expands with knowledge, models, OpenShell and blueprints so enterprises can govern AI agents on company data while keeping data in place.
Tags: NVIDIA Agent Toolkit, enterprise AI agents, data governance, OpenShell, Nemotron
Featured image: abstract, friendly illustration style showing connected data blocks, agent pathways and a secure runtime boundary, with no real people, no logos and no text.
The NVIDIA Agent Toolkit is for enterprise technology teams and agent builders who want AI agents to use company data while operating inside controlled boundaries. NVIDIA has expanded the open source NVIDIA Agent Toolkit with new tools so organizations can customize agents, ground them in private data, monitor their actions and set rules for how they operate. The toolkit brings together five building blocks: the new NVIDIA Agent Knowledge Foundation with NVIDIA CUDA-X libraries, NVIDIA open models including Nemotron Retriever and the new Kumo family, NVIDIA OpenShell secure agent runtime, NVIDIA NemoClaw secure agent blueprints, and tools and skills such as NeMo Relay and NeMo Compass. The design follows a production need: agents must combine models, data, runtime, tools and a supporting harness while staying trustworthy from development through deployment.
Making company data useful to agents
The NVIDIA Agent Knowledge Foundation helps agents retrieve the right company information, understand what it means for the business, and use it to make predictions or guide next decisions. NVIDIA positions the foundation so companies can adopt it without moving data, rebuilding existing workflows or changing governance controls.
The foundation includes NVIDIA CUDA-X libraries: cuDF, cuGraph, cuML, cuOpt and cuVS. It also adds the new Auto-Ontology library, which converts raw company data into familiar business concepts such as customers, products and orders, then maps their relationships.
The model layer includes NVIDIA Nemotron Retriever and Kumo Structured open models: Kumo Tabular, Kumo Relational, Kumo Time Series and Kumo Forecast. They are designed for retrieval, question answering and prediction over enterprise data.
Keeping agents inside secure boundaries
OpenShell defines the secure runtime boundary. It lets enterprises decide which data and actions an agent may use. NemoClaw supplies reference designs for secure, specialized agents. The newest NemoClaw blueprint for agent optimization offers a step-by-step starting point for improving agent intelligence, making agents better at their jobs over time.
For readers, the runtime layer matters because useful agents are not enough; they also need limits. Open source components may help teams inspect and adapt the stack, while OpenShell and NemoClaw aim to make secure operation a design requirement rather than an afterthought.
Early uses and reported performance gains
The new NVIDIA Agent Knowledge Foundation is already appearing in enterprise deployments. NVIDIA's supply chain organization is using NVIDIA Agent Knowledge Foundation through Palantir Foundry, Ontology and AIP to deploy specialized agents, create actionable intelligence from proprietary data, predict outcomes and optimize decision-making.
Databricks combines NVIDIA Kumo Anomaly with Databricks Genie Ontology to enable context-aware agents. Snowflake customers can use NVIDIA Kumo Engine and Kumo Relational open models inside Snowflake Notebook environments for more predictive capabilities.
Nemotron Retriever models make search faster and more accurate. Amazon OpenSearch Service used the embedding model to improve relevance with fewer search steps and virtually no added latency. SAP's fine-tuned reranker ran nearly 3x faster. ServiceNow improved reranking quality while cutting latency by nearly a quarter. Zoom is using Nemotron Retriever models to strengthen document understanding, search relevance and context retrieval for Zoom AI Services.
Other companies building on NVIDIA Agent Knowledge Foundation include Amazon Web Services (AWS), ClickHouse, Cloudera, Cohesity, Dassault Systèmes, Dell, Elastic, Exa, IBM, NetApp, Oracle, Perplexity, Starburst and VAST Data.
What to watch next
NVIDIA points readers to its Europe announcements and a GTC Berlin session on building, deploying and optimizing specialized agents in production. The expanded toolkit points to a practical next step for enterprise agents: connect them to company data, then control them with a runtime and reusable blueprints. Teams already working with Palantir, Databricks, Snowflake, AWS or other named builders may see direct relevance, because the announcement focuses on using existing data and production workflows rather than replacing them.
Comments (0)