NVIDIA Expands Agent Toolkit With Kumo Models and Secure Enterprise Data Grounding
NVIDIA expands its open source Agent Toolkit with Kumo models, secure runtimes and data grounding, backed by Palantir, AWS, SAP and more enterprise users.
Summary
On October 21, 2026, NVIDIA expanded its open source Agent Toolkit to ground enterprise agents in company data while preserving existing data locations, workflows and governance. Agent Knowledge Foundation uses CUDA-X libraries cuDF, cuGraph, cuML, cuOpt, cuVS and Auto-Ontology to turn raw data into connected business concepts such as customers, products and orders, helping agents retrieve context, interpret it, predict outcomes and support decisions. Its open models include Nemotron Retriever, Kumo Structured, Kumo Tabular, Kumo Relational, Kumo Time Series and Kumo Forecast.
OpenShell provides runtime boundaries controlling agent access and actions. NemoClaw offers secure, specialized agent blueprints, including a new step by step optimization design, while NeMo Relay and NeMo Compass add tools and skills. Palantir Foundry, Ontology and AIP now support NVIDIA supply chain agents for proprietary data intelligence, predictions and optimization. Databricks pairs Kumo Anomaly with Genie Ontology, while Snowflake is enabling Kumo Engine and Kumo Relational in Snowflake Notebooks.
Nemotron Retriever improved Amazon OpenSearch Service relevance with fewer search steps and virtually no added latency. SAP’s fine-tuned reranker ran nearly 3x faster, while ServiceNow improved reranking quality and cut latency by nearly a quarter. Zoom is using the models for document understanding, search relevance and context retrieval. AWS, ClickHouse, Cloudera, Cohesity, Dassault Systèmes, Dell, Elastic, Exa, IBM, NetApp, Oracle, Perplexity, Starburst and VAST Data are also building on the foundation. A GTC Berlin session will cover building, deploying and optimizing specialized production agents.
Positives
- SAP’s fine-tuned Nemotron Retriever reranker ran nearly 3x faster.
- ServiceNow improved reranking quality while reducing latency by nearly a quarter.
- Amazon OpenSearch Service improved relevance with fewer search steps and virtually no added latency.
- Agent Knowledge Foundation works without moving company data, rebuilding workflows or changing governance controls.
- Palantir Foundry, Ontology and AIP are already supporting specialized agents within NVIDIA’s supply chain organization.
Risks & concerns
- Agent usefulness still depends on the quality of available company knowledge and the boundaries governing its actions.
- OpenShell must control what agents can access and do, making runtime enforcement central to enterprise trust.
- Published performance figures cover Amazon OpenSearch Service, SAP and ServiceNow, while most named integrations lack comparable metrics.
- Availability and rollout details for the new Kumo family and agent optimization blueprint remain limited.