Nvidia Expands Agent Toolkit With Data Grounding Tools

Nvidia used its GTC Berlin event on October 21, 2026 to announce an expansion of the Agent Toolkit, the open source stack it has been assembling for enterprises that want to build agents rather than buy a single chatbot. The headline addition is the Agent Knowledge Foundation, a layer meant to turn a company’s existing records into something an agent can query directly, using Nvidia’s CUDA-X libraries and a new Auto-Ontology library that maps raw data into concepts like customers, products and orders. Nvidia says enterprises can adopt the foundation without moving their data, rebuilding workflows or changing governance controls already in place.

The toolkit also gains new Kumo family models, Kumo Tabular, Kumo Relational, Kumo Time Series and Kumo Forecast, which sit alongside the existing Nemotron Retriever models for search and retrieval. The runtime side is unchanged in name but still central: OpenShell remains the secure boundary that controls what an agent can access and do, and NemoClaw supplies reference blueprints for building agents inside that boundary, including a newest blueprint aimed specifically at improving agent performance over time.

Nvidia named a wide set of partners already working with the Agent Knowledge Foundation, including Palantir, Databricks, Snowflake, Amazon Web Services, SAP, ServiceNow and Zoom, each wiring Nvidia’s retrieval or anomaly detection models into their own platforms. The breadth of that list is the real signal here: Nvidia is positioning itself not as a model vendor competing on one leaderboard, but as the layer that sits underneath several companies’ separate agent products at once.

The numbers

Nvidia’s own examples give a sense of what the retrieval layer is meant to buy a customer. Amazon OpenSearch Service reported improved relevance with fewer search steps and little added latency after adopting the Nemotron Retriever embedding model. SAP said its fine-tuned reranker ran nearly 3x faster using the same family of models, and ServiceNow reported better reranking quality while cutting latency by close to a quarter. None of these figures are independently audited, and Nvidia did not publish the underlying benchmarks or baseline conditions, so they describe vendor-reported gains rather than a controlled comparison across the stack.

For a reader deciding whether to adopt an agent framework built on a single vendor’s retrieval and runtime layer, the open source licensing matters as much as the feature list, since it is the difference between a reversible choice and a locked-in one.