NVIDIA releases Agent Toolkit combining Nemotron models, NemoClaw blueprints, and OpenShell runtime for enterprise AI agents

NVIDIA has published a comprehensive framework for enterprise AI agent development under the NVIDIA Agent Toolkit banner, consolidating its Nemotron open models, NemoClaw agentic blueprints, and OpenShell secure runtime into a single architecture for building customizable domain-specific agents.

What's new

The toolkit packages three components into a unified agent stack:

Models — Nemotron open models form the base. According to NVIDIA, "NVIDIA Nemotron open models give teams flexibility to customize, evaluate and deploy agents for their own needs." These can be fine-tuned for domain-specific tasks such as protein design, chip verification, or security alert triage.

Tools and skills — NemoClaw blueprints provide agent behavior patterns. NVIDIA describes them as offering "patterns for safer agent behavior, delivering accurate results at lower costs, with tools and skills connecting agents to concrete actions."

Runtime — OpenShell handles secure execution inside enterprise systems. "The NVIDIA OpenShell runtime helps agents operate safely inside the systems where work gets done," according to the announcement.

NVIDIA highlighted real-world deployments: CrowdStrike security agents built on the toolkit are achieving 98.5% accuracy in alert triage. Cadence and Synopsys are using it for autonomous chip design workflows. Life sciences teams are deploying domain models for protein design research.

Context

NVIDIA has been steadily building out its software layer above the GPU. The Agent Toolkit consolidates product lines that have arrived at different intervals: Nemotron model weights were open-sourced earlier in 2026, NemoClaw blueprints were announced at DTW Ignite 2026 focused on telecom automation, and OpenShell has been part of the enterprise AI infrastructure push.

This announcement frames those components as a single enterprise platform, targeting the gap between generic LLM APIs and production-ready vertical agents that organizations can own, customize, and control.

Why it matters

Enterprise AI adoption has increasingly stalled on the question of customization and trust: general-purpose models handle breadth but organizations need agents they can tune, evaluate, and constrain to a domain. The toolkit's three-layer design — customizable models, safety-pattern blueprints, and a controlled runtime — is a direct response to that concern.

The open-weights Nemotron foundation also means teams are not locked into NVIDIA's inference cloud; the same models can run on-premises via NIM or on any NVIDIA GPU deployment. For industries like healthcare, financial services, and semiconductors — where proprietary processes cannot be exposed to third-party APIs — that flexibility is a prerequisite, not a nice-to-have.