NVIDIA moves to seize runtime control for agent and autonomous workflows

Nvidia used GTC 2026 to push past chips and into the software layer that runs enterprise AI agents. (investor.nvidia.com) On March 16, Nvidia announced the open-source NeMo Agent Toolkit and the OpenShell runtime for “autonomous, self-evolving” enterprise agents. The company said the software is meant to raise agent safety, security and efficiency. (investor.nvidia.com) (developer.nvidia.com) The basic problem is that AI agents are not just models. They are chains of tools, memory, permissions and handoffs that need to be observed, tested and locked down while they run. (developer.nvidia.com 1) (developer.nvidia.com 2) Nvidia’s pitch is that NeMo Agent Toolkit sits across those moving parts, even when developers use other frameworks such as LangChain, LlamaIndex, Semantic Kernel and CrewAI. Nvidia says the toolkit adds profiling, observability and optimization across connected agent systems. (developer.nvidia.com 1) (developer.nvidia.com 2) OpenShell is the second half of the move. Nvidia describes it as an open-source runtime that provides a secure environment for running autonomous agents and open models such as Nemotron. (investor.nvidia.com) (nvidianews.nvidia.com) Nvidia said 17 software platforms were advancing enterprise and physical AI agents with the new stack at launch. The list included Adobe, Atlassian, Amdocs, Box, Cadence, Cisco, Cohesity, CrowdStrike, Dassault Systèmes, IQVIA, Red Hat, SAP, Salesforce, Siemens, ServiceNow and Synopsys. (nvidianews.nvidia.com) (venturebeat.com) Adobe separately said it would use Nvidia Agent Toolkit and OpenShell in agentic creative and marketing workflows. Nvidia said Adobe’s stack would combine the toolkit with secure infrastructure and policy-driven controls. (nvidianews.nvidia.com) (blogs.nvidia.com) Nvidia tied the software push to a broader platform strategy at GTC. The same event featured Nemotron models, the AI-Q agentic search blueprint and Vera Rubin systems aimed at the next generation of agentic AI infrastructure. (nvidianews.nvidia.com 1) (nvidianews.nvidia.com 2) That package matters for one reason: the company is trying to define how agents run, not just what chips they run on. If customers adopt Nvidia’s toolkit, runtime and model stack together, Nvidia gains influence over orchestration, monitoring and security policies across the AI workflow. (developer.nvidia.com) (investor.nvidia.com) The company’s own language reflects that shift. Nvidia said the software is meant to speed “a generational shift in software and knowledge work,” putting the runtime layer alongside its hardware as part of the platform sale. (investor.nvidia.com) The near-term test is whether those launch partners move from announcements to production deployments. Nvidia has already made clear that, in the agent era, it wants to own the operating layer as well as the compute underneath it. (venturebeat.com) (nvidianews.nvidia.com)