THE APEX TIMES
NVIDIA expands its Agent Toolkit aimed at keeping enterprise AI agents grounded in company data
The company says new components for turning raw business information into “company intelligence,” plus a secure runtime boundary, are designed to help agents operate within governed limits.
NVIDIA on Tuesday expanded its open-source approach for building and operating AI “agents” in enterprise settings, focusing on a central problem: agents can only be useful if they have reliable access to the right data and if their actions are constrained by trustworthy boundaries.
In a post describing its latest enterprise platform work, NVIDIA framed agent development as a full stack exercise, bringing together models, runtime, data, tools and governance. The company said enterprises are using the open-source NVIDIA Agent Toolkit to customize agents, use company data to inform their answers and actions, monitor what agents do, and govern how they operate once deployed.
At the center of the expansion is what NVIDIA calls the NVIDIA Agent Knowledge Foundation. The foundation is positioned as a way to convert enterprise data into “company intelligence” that agents can use for tasks like finding relevant information, interpreting it in business terms, and making predictions that help people decide next steps.
NVIDIA said the foundation can be used without moving data, rebuilding existing workflows, or changing established governance controls. It also said the foundation includes NVIDIA CUDA-X libraries, naming cuDF (data frame acceleration), cuGraph (graph processing), cuML (machine learning), cuOpt (optimization), and cuVS, along with a new Auto-Ontology library intended to translate raw company data into familiar business concepts such as customers, products and orders, and map how those concepts relate to one another.
For retrieval and prediction, the foundation includes NVIDIA Nemotron Retriever and Kumo structured open models, which NVIDIA said help agents locate relevant information to answer questions and generate predictions. The post lists Kumo Tabular, Kumo Relational, Kumo Time Series and Kumo Forecast, tying them to different types of business data and analytic needs.
On the security and control side, NVIDIA said OpenShell provides the “secure runtime boundary,” giving enterprises the ability to control what agents can access and what they can do. NVIDIA also pointed to NVIDIA NemoClaw blueprints as reference designs for building specialized agents, and introduced a newest NemoClaw blueprint focused on agent optimization, described as a step-by-step starting point to improve an agent’s intelligence over time.
NVIDIA said the expanded Agent Knowledge Foundation is already being used by organizations across multiple industries, including examples of supply chain, data and cloud platforms. Palantir, NVIDIA’s supply chain organization, is using 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.
Other deployments cited in the post span data platforms and cloud services. Databricks was said to pair NVIDIA Kumo Anomaly with Databricks Genie Ontology to support context-aware agents, while Snowflake was said to enable customers to use NVIDIA Kumo Engine and Kumo Relational open models within Snowflake Notebook environments to unlock more predictive capabilities.
NVIDIA also highlighted performance-oriented outcomes from Nemotron Retriever models, stating that Amazon OpenSearch Service improved relevance with fewer search steps and virtually no added latency after incorporating an embedding model. SAP said its fine-tuned reranker ran nearly 3x faster, ServiceNow reported improved reranking quality while cutting latency by nearly a quarter, and Zoom said it is using Nemotron Retriever models to improve document understanding, search relevance and context retrieval for Zoom AI Services.
NVIDIA added that other companies are building on the Agent Knowledge Foundation, naming AWS, ClickHouse, Cloudera, Cohesity, Dassault Systèmes, Dell, Elastic, Exa, IBM, NetApp, Oracle, Perplexity, Starburst and VAST Data. The company also invited builders to join a GTC Berlin session focused on building, deploying and optimizing specialized agents in production.
Why It Matters
- Enterprises are increasingly looking for AI systems that do more than answer questions, and grounding those agents in company data is a practical bottleneck that vendors are trying to solve.
- By packaging retrieval, data-to-ontology mapping, prediction models and a runtime boundary together, NVIDIA is positioning its toolkit as an agent operating layer rather than a standalone model.
- The named integrations suggest NVIDIA is trying to make agent building portable across data platforms and enterprise software ecosystems, which could reduce deployment friction.
- If the cited performance and latency improvements hold broadly, they could make agent-driven search and decision support more usable in production settings.
Sources
Key Facts
- NVIDIA said its expanded open-source NVIDIA Agent Toolkit targets enterprise agent use by emphasizing trustworthy boundaries, governed access to data, and monitoring of agent behavior.
- The new NVIDIA Agent Knowledge Foundation is described as a way to turn company data into “company intelligence” agents can use for information lookup, interpretation in business terms, and prediction.
- NVIDIA cited components including CUDA-X libraries (cuDF, cuGraph, cuML, cuOpt and cuVS), an Auto-Ontology library to map raw data into business concepts, and Nemotron Retriever plus Kumo structured open models (Kumo Tabular, Kumo Relational, Kumo Time Series and Kumo Forecast).
- For runtime control, NVIDIA said OpenShell provides a secure runtime boundary, and it referenced NemoClaw blueprints including a new blueprint for agent optimization.
- The post listed early adopters or pilots including Palantir (via Foundry, Ontology and AIP), Databricks, and Snowflake, plus performance claims tied to Nemotron Retriever in products from Amazon OpenSearch Service, SAP, ServiceNow and Zoom.
Technology Related
Oracle rolls out ‘AI Agent Studio for Industry Applications’ aimed at speeding deployment of industry-focused AI agents
The new platform is positioned as a no-code and low-code environment for building, testing, orchestrating, deploying and managing AI agents and workflows tailored to specific business needs.
NVIDIA expands confidential computing, adding “Confidential Bare Metal” for developers running sensitive AI workloads on third-party infrastructure
The company says its updated security controls and new deployment option are designed to let frontier model builders keep control of models and data while scaling training and inference across global clouds and bare-metal environments.
NVIDIA pushes an ‘always-on’ factory-agent blueprint across Europe’s manufacturing ecosystem
A new early-access factory operations blueprint on GitHub is aimed at helping industrial developers deploy coordinated software agents across machines, inspection systems, and enterprise tools, using NVIDIA’s simulation, digital-twin, and frontier-model stack.
NVIDIA makes cuDSS generally available, aiming to speed up GPU-accelerated engineering and science simulations
The company says its cuDSS software library, part of the CUDA-X ecosystem, moves a key simulation step to GPUs to reduce time-to-results and expand what teams can model.
NVIDIA highlights expanding European “AI cloud” network, pushing open models and agent safety controls
NVIDIA says cloud partners across Europe, the Nordics, Africa, and Central Asia are adding regional compute capacity for training and inference, while adopting a reference architecture meant to constrain what AI agents can access and do.
NVIDIA’s GTC Berlin pitch: local “Spark” systems bring AI agents closer to where work happens
At GTC Berlin, NVIDIA highlighted a new lineup of local AI hardware and software, positioning “local intelligence” as the next platform for personal agents, developer workflows, and enterprise deployments that keep data on device.
Google Cloud and NVIDIA set general availability for A5X, a Vera Rubin-based platform for training and agentic reinforcement learning
The companies say NVIDIA Vera Rubin infrastructure delivered as integrated A5X instances will reach general availability next month, with planned expansion to Europe and a stated capacity path to as many as 960,000 Rubin GPUs.
Meta faces a profitability test as investors scan whether capex is paying off
After Meta reports earnings, attention is focused less on day-to-day revenue momentum and more on whether rising spending is translating into durable returns.
Meta slips ahead of earnings, while SpaceX-related chatter lifts Tesla and other megacap stocks, market roundup says
A market-focused roundup highlighted Meta’s slide as investors await the Facebook parent’s results, alongside a late-market surge tied to new comments from Elon Musk about exploring a Tesla merger.
NVIDIA expands Isaac open robotics platform with new humanoid reference design, datasets, and simulation pipelines
The company says its Isaac platform for building robotics systems is adding a reference humanoid plus accompanying frameworks, runtime elements, simulation workflows, and data intended to help developers train and test general-purpose robot behaviors.