On October 21, 2026, NVIDIA expanded its enterprise AI agent development platform, NVIDIA Agent Toolkit, and announced Agent Knowledge Foundation, which links internal company data to answers and predictions. It combines document-retrieval models, a library that organizes business concepts such as customers and orders, and Kumo, which makes predictions from tabular data. The setup is designed to connect everything from searching internal information to supporting decisions, while making use of existing data infrastructure. However, companies that adopt it still have to decide how to define the meaning of their data and which actions to base on the predictions.
A new foundation connecting retrieval, prediction and optimization
Agent Knowledge Foundation is a platform that combines models with data-processing libraries. NVIDIA says it can be used without moving data, rebuilding business processes or changing existing controls. The developer page likewise says it can be used alongside existing data platforms and agent development environments, and that components can be adopted individually.
The difference between the components lies in what they extract from internal data. Retrieval fetches relevant information, prediction estimates future outcomes and risks, and optimization allocates resources within constraints. Matching the process descriptions on NVIDIA's developer page with each product's functions gives the following breakdown.
| Role in the business | Main components | What it handles |
|---|---|---|
| Find relevant information | Nemotron Retriever, cuVS | Reads document content, searches for relevant information and ranks supporting evidence |
| Define business meaning | Auto-Ontology | Drafts definitions of concepts and relationships such as customers, products and orders |
| Prepare data and examine relationships | cuDF, cuGraph | Joins and transforms tables and analyzes connections |
| Predict outcomes and risks | Kumo, cuML | Makes predictions with pretrained models and, depending on the use case, builds custom machine-learning processing |
| Allocate within constraints | cuOpt | Optimizes routes, schedules and resource allocation according to objectives and constraints |
The source is the process descriptions and product entries on the Agent Knowledge Foundation developer page. The table organizes features as of October 22, 2026 by role; it is not a performance ranking or an adoption sequence for all products.
For example, forecasting demand is a different process from deciding allocation while accounting for inventory and supply constraints. Even if Kumo produces a demand outlook, that figure does not automatically determine which supplier to order from or how much. Adding objectives and constraints to a prediction is what allows the process to move on to optimization.
The new component that sets up that groundwork is Auto-Ontology. An ontology is a set of definitions that represents business concepts such as customers and orders, along with how they relate to one another. The library creates draft definitions as a starting point from a company's data and documents, and business experts then review, revise and certify them. The design separates an AI's ability to read data from the confirmation of what that data means in the business.
Kumo is chosen according to the shape of the data
Kumo is a family of open models and libraries that make predictions from structured data. The model to choose differs depending on whether you are working with a single table or with relationships spanning multiple tables. The Kumo developer page describes the use cases as follows.
| Model | Main data | What it predicts or detects |
|---|---|---|
| Kumo Tabular | A single table | Classification and regression that use labeled examples as context |
| Kumo Relational | Multiple linked tables | Classification and regression that use relationships between records |
| Kumo Forecast | Time series of past numerical values | Future values such as demand or throughput |
| Kumo Anomaly | Time series | Unusual movements and signs of anomalies |
This table is based on the use-case descriptions on the Kumo developer page. The October 21 announcement blog also lists "Kumo Time Series," so the names do not match exactly. This should not be read to mean that every model in the table made its debut on the same day.
Tabular places labeled examples in the input context and is described as producing results without training a model for each prediction task. Relational also uses labeled examples and the relationships between tables, and is intended for tasks such as predicting customer churn. Forecast, on the other hand, predicts the future from past numbers, while Anomaly detects anomalies in time series, so even with time-series data the questions they answer differ.
The software used to run them also differs. Relational and Tabular use "Kumo Structured Data Models," while Forecast and Anomaly use "Kumo-TS." Kumo Engine is also provided for building, training and evaluating custom relational-data prediction models. Even where a model can skip per-task training, deployment preparation does not disappear. Before adopting Kumo Engine, NVIDIA advises setting up a Python environment, local storage and data access.
From the earlier Palantir collaboration to existing platforms
NVIDIA also said its own supply chain organization is using Agent Knowledge Foundation through Palantir Foundry, Ontology and AIP. This effort has a precursor. In an announcement on September 10, 2026, the two companies disclosed a collaboration that combines Nemotron with Palantir's platform, deploying it first in NVIDIA's own supply chain.
The September description cited, as a starting point, decisions about allocating components, with uses such as spotting supply constraints early and evaluating alternatives. A tuned Nemotron model explains recommended actions and the trade-offs of allocations, while cuOpt handles constraints and allocation scenarios. The design leaves the final decision to supply chain experts. The latest announcement positions this earlier collaboration as an example of use on an enterprise data platform.
Combinations with other data platforms were also shown. NVIDIA says Databricks combines Kumo Anomaly with Genie Ontology, and Snowflake lets users use Kumo Engine and Kumo Relational in its notebook environment. The features and execution environments each company handles are not the same, and the number of adopting companies alone does not indicate that all features have reached a stage where they can be used uniformly.
Definitions, evaluation and permissions remain for adopting companies to decide
What Auto-Ontology generates are draft definitions for business experts to check. This step is consistent with the explanation that data can be used where it already sits. Even if it reduces the effort of migration, it does not automatically settle what counts as the same customer or order.
Validation on the adopter's side can be thought of in stages: from business definitions to prediction to subsequent action. First, experts confirm whether the definitions of concepts and relationships fit the actual business. Next, the predictions need to be evaluated on the company's own data and objectives, and the company must decide which decisions the results can inform. For example, producing a forecast of rising demand and authorizing orders based on that forecast are separate decisions.
Within Agent Toolkit, OpenShell serves as the runtime environment that controls what agents can access and do, and NemoClaw provides a reference design for building agents for specific tasks. The structure combines components that give data meaning and make predictions with components that control the range of actions. However, this announcement did not specify a general availability date, pricing or Japanese-language support conditions common to all features, so terms of use need to be checked for each component and integration partner chosen.
For companies, the adoption decision comes down to where on their existing platform to add retrieval and prediction, and which decisions to leave to people. If they can confirm business definitions, evaluate predictions and set the scope of actions, a path opens to take internal data beyond search and tie it to decisions on allocation and planning.