NVIDIA expands Agent Toolkit: enterprise agents with your data and a secure runtime

What it's about

On October 21, 2026, NVIDIA expanded its open-source Agent Toolkit. An AI agent is a program that doesn't just answer questions but carries out tasks on its own, such as querying data or triggering actions. NVIDIA's rationale for the expansion: agents are only as useful as what they know, and only as trustworthy as the boundaries they operate within. The toolkit is meant to help companies customize agents, feed them company data, monitor them and keep them under control.

What the toolkit contains

According to NVIDIA, the toolkit has five parts:

• Agent Knowledge Foundation (new), with CUDA-X libraries, meaning GPU-accelerated software building blocks from NVIDIA
• Open models, including Nemotron Retriever and the new Kumo family
• OpenShell, a secure runtime for agents
• NemoClaw, blueprints (reference designs) for secure, specialized agents
• Tools and skills such as NeMo Relay and NeMo Compass

NVIDIA argues that anyone who wants to build agents that do real work needs an open stack of models, harness (the control layer around the model), runtime, data and tools.

Agent Knowledge Foundation: making company data usable

The Foundation is meant to turn raw data into knowledge agents can use: find relevant information, understand what it means for the business and derive predictions from it. According to NVIDIA, this works without moving data, rebuilding existing workflows or changing governance rules.

It includes the CUDA-X libraries cuDF (dataframes), cuGraph (graphs), cuML (machine learning), cuOpt (optimization) and cuVS (vector search). New is the Auto-Ontology library. An ontology is a structured description of concepts and their relationships. The library translates raw data into business concepts such as customers, products and orders, and maps how they relate to each other.

Also included are the open models Nemotron Retriever and the Kumo Structured models: Kumo Tabular, Kumo Relational, Kumo Time Series and Kumo Forecast. They help agents find the right company information and make predictions. The sources don't give parameter counts, licenses or download locations.

Security and optimization

OpenShell draws the line on what agents can access and do. The NemoClaw blueprints provide templates for specialized, secure agents. The newest blueprint, for agent optimization, is meant to offer a step-by-step entry point for making agents better over time.

Who's using it (claims from NVIDIA)

Other companies building on the Foundation, according to NVIDIA, include AWS, ClickHouse, Cloudera, Cohesity, Dassault Systèmes, Dell, Elastic, Exa, IBM, NetApp, Oracle, Perplexity, Starburst and VAST Data. A session at GTC Berlin covers the topic.

Our take

The announcement shows NVIDIA's strategy of offering not just chips but a full software stack for enterprise agents. A blockchain.news report places it in a series of earlier expansions in 2026, such as BioNeMo, Omniverse and PhysicsNeMo. The benefits are mainly for companies that want to run agents on their own data and value control. The caveat: all performance figures and deployment examples come from NVIDIA itself and haven't been independently verified. It's also unclear how openly the individual components are licensed and how tightly the stack ties you to NVIDIA hardware, since the CUDA-X libraries are built for NVIDIA GPUs.