NVIDIA expands Agent Toolkit with a foundation for enterprise knowledge and predictions

NVIDIA expanded its Agent Toolkit on October 21 with Agent Knowledge Foundation, a collection of models and libraries designed to connect enterprise agents with company information. The announcement shifts attention from a model’s general knowledge to the documents, tables and business relationships needed for specific decisions. NVIDIA places the new foundation alongside its OpenShell runtime, NemoClaw blueprints and tools for building and improving specialized agents.

The company described deployments spanning data platforms rather than a single new application. NVIDIA’s supply-chain organization is working with Palantir’s Foundry, Ontology and AIP, while Snowflake is bringing Kumo capabilities into notebook environments. Other partners are applying Nemotron Retriever models to search and document understanding. These examples illustrate different integration points; the announcement does not mean every component is automatically available to every customer of each named platform.

The developer material presents the foundation as modular. Nemotron Retriever supplies document processing, embedding and reranking, while CUDA-X libraries support data preparation, similarity search, graph analysis, machine learning and optimization. Existing databases, ontologies and agent frameworks can remain part of the workflow. The individual pieces can be adopted separately, allowing an application to improve one stage of its information pipeline without replacing the entire enterprise data environment.

Auto Ontology addresses the relationships behind that information. It generates an initial representation of business concepts from enterprise data and documentation, which business experts can test, refine and certify. This adds a layer between raw records and an agent’s response: a customer, order or product can be understood through its connections to other records. The generated starting point still requires business validation rather than being treated as an authoritative description by default.

Kumo adds prediction over structured data. NVIDIA describes a family of open models, libraries and tools that work with tables, relationships between tables and time-series signals. The intended tasks include forecasting outcomes, assessing risk, detecting unusual activity and recommending actions. Such tasks differ from finding a paragraph in a document: the output concerns a possible future outcome or a pattern in records, rather than simply retrieving an existing statement.

The approach uses pretrained models to reduce work repeated across conventional predictive machine-learning projects. Those projects often require separate data preparation, feature engineering, training, tuning and deployment for each task. Kumo aims to make more of that capability reusable across structured-data problems. The developer description therefore expands the role of an enterprise agent beyond language generation, while still leaving the application’s data, question and evaluation central to whether a prediction is useful.

Runtime controls form another part of the stack. OpenShell’s documentation describes policies governing readable and writable paths, process identities, and which network destinations a program can reach. Access that a policy does not allow is denied. Some controls are fixed when a sandbox starts, while network rules can change during operation. These mechanisms concern what agent software is permitted to do, a separate question from whether its retrieved evidence or prediction is correct.

At the point where a prediction has to become an operational decision, NVIDIA pairs these components with cuOpt. The optimization library works with objectives and constraints to produce routes, schedules and resource allocations. cuGraph supplies analysis of connected records, while cuML supports custom machine-learning pipelines. The developer material treats these as independently usable components, giving teams several ways to connect the new models with existing analytical systems.