NVIDIA Builds an Agent Software Stack on Top of Its Chips

NVIDIA published a description of its Agent Toolkit on June 23, moving explicitly into the enterprise agent software market rather than waiting for third parties to fill that role above its hardware.

The toolkit bundles four components: the Nemotron family of open-weight models that teams can fine-tune and self-host, NemoClaw blueprints that define patterns for tool use and agent behavior, the OpenShell runtime that executes agent workflows, and a tool and skill layer that connects agents to external systems. The combination gives enterprises a single integrated stack they can deploy without assembling parts from separate vendors.

NVIDIA’s own blog post announced the toolkit. The post describes acceleration and accuracy gains in general terms, but the claimed figures come from NVIDIA itself, not from independent audits. One specific number is attributed to a named customer: CrowdStrike, the cybersecurity firm, is cited as running security-alert triage agents that reach 98.5 percent accuracy. That figure is CrowdStrike’s own, sourced through NVIDIA’s announcement. Cadence and Synopsys, two electronic design automation companies, are listed as using the toolkit for autonomous chip design workflows. Palantir, SAP, ServiceNow, Siemens, and Dassault Systemes are named as embedding agent capabilities into their enterprise platforms via the toolkit.

The move is structurally significant because it extends NVIDIA’s position up the value chain. Selling GPUs places NVIDIA at the infrastructure layer. Shipping an opinionated agent runtime means NVIDIA now has a stake in how enterprise workloads are designed, not only on what hardware they run. Companies that standardize on Nemotron models, NemoClaw blueprints, and OpenShell face the same switching costs that any platform lock-in creates: fine-tuned weights, behavioral presets, and deployment pipelines built around one vendor’s abstractions.

The OpenShell runtime is the component worth watching most closely on that dimension. A runtime that controls how agents access tools, manage state, and call external systems becomes the operating substrate for agent workloads. That is a durable position if enterprises standardize on it, and a difficult one to displace later. The announcement does not include specifics on how OpenShell isolates agent actions or enforces permission boundaries, so the “secure” framing rests on NVIDIA’s own characterization for now.

The toolkit supports third-party orchestration frameworks, including Hermes Agents and OpenClaw, which limits some concerns about forced monoculture. Teams can bring their own orchestrator and use the Nemotron models and NemoClaw blueprints independently. That design also gives NVIDIA a softer adoption path: enterprises already running LangChain or another framework can adopt pieces without replacing what they have.

The BioNeMo Toolkit is positioned separately for life sciences, allowing agents to call domain models for protein design, genomics analysis, and biomarker discovery. The announcement says this compresses research workflows from months to days. That claim is NVIDIA’s own.

Enterprise AI teams evaluating agent infrastructure in the second half of 2026 now have NVIDIA competing directly with agent platform vendors; any evaluation that previously excluded NVIDIA on the grounds that it was a hardware company needs to be updated.

Source: NVIDIA blog post by Justin Boitano, published June 23, 2026, at blogs.nvidia.com.