Enterprise AI agents are failing in production, and it is not a model problem. Forty-six percent of organizations cite integration with existing systems as their top agent blocker, and most teams that get past the demo stage run straight into the same wall: the agent is connected to your database but has no idea what your data means. At GTC Berlin this week, NVIDIA expanded its Agent Toolkit with the Agent Knowledge Foundation — an open-source collection that gives agents genuine business context without moving data or bypassing access controls.
The failure mode is subtle enough that teams often miss it until they are deep in production. You point an agent at Snowflake, it retrieves records, and it still tells Finance the wrong revenue number because it resolved “customer” using Marketing’s definition. RAG fixes the unstructured-text retrieval problem. It does not fix the authority problem — which of three competing definitions of “customer” is the canonical one. NVIDIA’s new Auto-Ontology library does: it builds a knowledge graph from your existing metadata, maps business concepts like customers, products, and orders across systems, and exposes that semantic layer through MCP, CLI, and API. The data stays put. Governance stays intact. The agent finally understands what it is looking at.
Kumo Brings Prediction Models to Structured Data #
Most enterprise data is not text — it is tables, relational records, and time series. That is exactly where traditional LLM-based retrieval breaks down, and exactly where the Kumo model family picks up. NVIDIA acquired Kumo in June 2026 for more than $400 million for its foundation models trained on structured data, and those models are now open and part of the Agent Toolkit. The Kumo family ships four open models: Kumo Tabular, Kumo Relational, Kumo Time Series, and Kumo Forecast. They handle what structured-data agents actually need — forecasting outcomes, scoring risk, detecting anomalies, and recommending actions — all from the tables you already have.
Kumo Tabular landed on Hugging Face in September 2026 and is available right now. Databricks is pairing Kumo Anomaly with Databricks Genie Ontology. Snowflake customers get Kumo Relational natively inside Snowflake Notebooks. If your data lives in either platform, the integration path is already there — no migration required. More details on the full Kumo lineup are on the NVIDIA Kumo developer page.
OpenShell Is the Part You Can Use Today #
While Kumo and Auto-Ontology are the long-term bets, OpenShell is what you can actually adopt this afternoon. It is an Apache 2.0 secure runtime sandbox for autonomous agents, and the install is two commands:
curl -LsSf https://raw.githubusercontent.com/NVIDIA/OpenShell/main/install.sh | sh
openshell sandbox create
OpenShell runs with deny-by-default policies, live policy updates, and a full audit trail. Agents get exactly the permissions they need and nothing more. If you have engineers running Claude Code or Codex on your infrastructure with real credentials, OpenShell is the missing control layer. It scales from a single developer machine to an enterprise GPU cluster using the same primitives — the sandbox you test locally is the sandbox you run in production. The OpenShell repository is on GitHub under Apache 2.0.
Enterprise Adoption Is Already in Motion #
NVIDIA launched Agent Knowledge Foundation with 17 enterprise partners committed: Adobe, Salesforce, SAP, ServiceNow, Siemens, CrowdStrike, Atlassian, Cadence, Synopsys, IQVIA, Palantir, Box, Cohesity, Dassault Systèmes, Red Hat, Cisco, and Amdocs. LangChain — with over one billion downloads — is integrating Agent Toolkit components into its library. This is not a roadmap announcement. Production integrations are live or imminent. The full official breakdown is on the NVIDIA blog.
One important housekeeping note: the older NeMo Agent Toolkit is being deprecated and reaches end of life at the end of 2026. If your stack depends on it, NVIDIA’s recommended migration paths are NeMo Relay for observability and governance, NeMo Helix for evaluation and optimization, and NeMo Fabric for harness interfaces. Plan accordingly.
The Bigger Picture #
NVIDIA is making a platform bet that extends well beyond GPU sales. Agent Knowledge Foundation is the latest layer of a stack that now runs from Vera Rubin silicon to agent runtime sandboxes to structured-data prediction models. The company is positioning itself as infrastructure for enterprise agent deployment regardless of which model or orchestration framework sits on top. If the data grounding layer, the safety layer, and the prediction layer all default to NVIDIA tooling, NVIDIA owns the enterprise agent stack even when organizations run OpenAI, Anthropic, or open-weight models. The independent analysis from The New Stack frames this well.
For developers building agents today, the practical upshot is clear. OpenShell is production-ready and free. Kumo Tabular is on Hugging Face now. Auto-Ontology is open source and MCP-compatible. The pieces are there. The question is whether your organization has the data governance posture to use them — and for the 58% of enterprises that have cited data readiness as their primary blocker five years running, NVIDIA just shipped the shovel.