NVIDIA open‑sources agent platform
NVIDIA has open‑sourced an enterprise agent platform that 17 major firms have already adopted as an alternative to proprietary stacks. The list of early adopters includes Adobe, Salesforce, SAP and Cisco — a signal that big enterprises want an agent layer they can run and inspect themselves while still leaning on NVIDIA’s GPU advantage. (x.com)
NVIDIA is trying to stop enterprises from building AI agents as black boxes they rent from someone else. On March 16, 2026, it released an open agent stack called NVIDIA Agent Toolkit, and 17 software companies signed on at launch, including Adobe, Salesforce, SAP, Cisco, ServiceNow, Siemens, Red Hat, and Atlassian. (nvidianews.nvidia.com) An AI agent is software that does jobs in steps instead of just answering one prompt. NVIDIA says its toolkit is for agents that can choose data sources, use tools, and complete tasks inside company systems with built-in evaluation showing how each answer was produced. (nvidianews.nvidia.com) The new part is not just “another model.” NVIDIA bundled open models, an open search-and-reasoning blueprint called AI-Q, and an open runtime called OpenShell, which is the layer that actually runs the agent while enforcing security and privacy rules. (investor.nvidia.com) OpenShell is the piece aimed at the fear every information-technology team has: an agent that can take actions inside real systems without guardrails. NVIDIA says OpenShell adds policy-based security, network controls, and privacy guardrails so companies can decide what an agent is allowed to touch before it starts clicking around their software. (investor.nvidia.com) The other enterprise problem is that agents are expensive and hard to debug. NVIDIA’s NeMo Agent Toolkit tracks tool usage, coordination between agents, latency, and cost, and it exports those traces to observability systems so teams can see where an agent got slow, wrong, or expensive. (developer.nvidia.com) That matters because most companies do not want one giant agent doing everything. NVIDIA’s own developer docs describe teams using different agent frameworks together, with one workflow builder connecting them and one monitoring layer measuring them, which is closer to how large companies already run databases, cloud services, and internal tools. (developer.nvidia.com) NVIDIA is also selling a cost story, not just a control story. The company says its AI-Q setup uses frontier models for orchestration and its own open Nemotron models for research, and that hybrid mix can cut query costs by more than 50 percent while keeping top scores on DeepResearch Bench and DeepResearch Bench II. (nvidianews.nvidia.com) The partner list shows where NVIDIA wants to sit in the stack. Adobe makes creative and document software, Salesforce runs customer data, SAP runs finance and supply chains, and Cisco runs networks, so if all four build on the same agent layer, NVIDIA is no longer just selling graphics processing units underneath the system; it is shaping how the software above them behaves. (nvidianews.nvidia.com) NVIDIA’s product pages make the pitch even clearer: NeMo supports cloud, on-premises, and hybrid deployment, and it works with open and proprietary models. That gives companies a way to keep agents inside their own infrastructure while still using NVIDIA chips and software to train, tune, and run them. (nvidia.com) So this launch is really a bid to make NVIDIA the plumbing for enterprise AI work. If software companies adopt the toolkit and customers keep demanding agents they can inspect, benchmark, and lock down themselves, the valuable layer may shift from the chatbot on top to the runtime, monitoring, and model-routing system underneath. (nvidianews.nvidia.com)
