NVIDIA targets enterprise software with a “full-stack” agent architecture

NVIDIA is repositioning itself beyond infrastructure and into enterprise software, introducing a “full-stack” agent architecture designed to move AI from generation to execution. The company’s announcements outline a coordinated push across models, runtimes, orchestration layers and application integrations, signalling a change in how enterprise systems are built and operated.

At the centre of this strategy is the NVIDIA Agent Toolkit, which combines open models, agent frameworks and runtime environments to support autonomous systems that can determine and complete tasks. The approach reflects a broader transition in enterprise AI, where the focus is moving from outputs to workflows, and from tools to systems that can act with limited human intervention.

From models to autonomous execution

The Agent Toolkit introduces a modular stack that includes NVIDIA Nemotron models, the NVIDIA AI-Q agent blueprint (for deep research tasks), and the NVIDIA OpenShell runtime (an isolated sandbox for agent execution). Together, these components enable developers to build agents that can perceive, reason and act on enterprise data, while maintaining visibility into how decisions are made.

This design moves beyond traditional generative AI, where systems respond to prompts, toward autonomous execution across workflows. Agents can select data sources, determine the depth of analysis required and generate context-aware outputs, creating a more adaptive layer within enterprise software.

“Claude Code and OpenClaw have sparked the agent inflection point — extending AI beyond generation and reasoning into action,” said Jensen Huang, founder and CEO of NVIDIA. “Employees will be supercharged by teams of frontier, specialized and custom-built agents they deploy and manage. The enterprise software industry will evolve into specialized agentic platforms, and the IT industry is on the brink of its next great expansion.”

The change introduces new operational requirements, particularly around governance, traceability and system reliability. NVIDIA’s inclusion of a built-in evaluation system within AI-Q reflects these needs, allowing enterprises to understand how outputs are generated rather than treating them as opaque results.

Security and control become core to agent systems

As agents take on more responsibility, the underlying runtime environment becomes critical. NVIDIA OpenShell provides policy-based controls for security, network access and data privacy, defining how agents interact with enterprise systems and external services.

This layer addresses a central challenge in enterprise AI adoption, where autonomy must be balanced with control. By embedding guardrails at the runtime level, NVIDIA is positioning security and compliance as foundational to agent deployment rather than an afterthought.

NemoClaw extends this model into the OpenClaw ecosystem, simplifying deployment by packaging models and runtime components into a single command. The system allows agents to operate across local and cloud environments, using open models such as NVIDIA Nemotron alongside frontier models accessed through a privacy-controlled routing layer.

“OpenClaw opened the next frontier of AI to everyone and became the fastest-growing open source project in history,” said Jensen Huang. “Mac and Windows are the operating systems for the personal computer. OpenClaw is the operating system for personal AI. This is the moment the industry has been waiting for — the beginning of a new renaissance in software.”

The emphasis on always-on agents introduces new infrastructure considerations, particularly around dedicated compute and persistent execution. These systems are designed to run continuously, completing tasks and building tools over time rather than responding to isolated queries.

Ecosystem integrations push agents into enterprise workflows

NVIDIA’s strategy extends beyond core tooling into partnerships with major enterprise software providers. The company identified 17 inaugural partners, including Adobe, Salesforce, and SAP, that have announced plans to integrate Agent Toolkit components to expand the capabilities of their platforms.

These integrations position agents as active participants in business workflows rather than peripheral assistants. For example, agents embedded within enterprise applications can monitor processes, automate decisions and execute tasks across systems, reducing the need for manual coordination.

Adobe’s collaboration with NVIDIA illustrates this direction, combining agentic frameworks with creative and marketing workflows. The partnership integrates NVIDIA models, runtimes and infrastructure into Adobe’s platforms, enabling long-running, personalised and cost-efficient agent-driven processes across content creation and campaign execution.

“Content creation is exploding, and our partnership with NVIDIA is grounded in a shared vision to reinvent creative and marketing workflows with the power of AI,” said Shantanu Narayen, chair and CEO of Adobe. “As AI transforms how marketing teams and media and entertainment studios work, Adobe and NVIDIA will bring together our Firefly models, CUDA libraries into our applications, 3D digital twins for marketing, and Agent Toolkit and Nemotron to our agentic frameworks to deliver high-quality, controllable and enterprise-grade AI workflows of the future.”

Across industries, these integrations reflect a shift toward domain-specific agents tailored to particular functions, from engineering design and customer operations to content production and data management. The result is a fragmented but specialised ecosystem of agentic systems embedded within enterprise platforms.

Orchestration and inference shape the economics of agents

As agent systems scale, orchestration and inference become central to performance and cost. NVIDIA Dynamo 1.0 introduces a distributed inference layer designed to manage GPU and memory resources across workloads, functioning as an operating system for AI factories.

The system routes requests, manages memory and balances workloads across clusters, enabling more efficient execution of long-running and multi-step tasks. This is particularly relevant for agentic systems, where workflows may involve sustained reasoning and interaction with multiple data sources.

“Inference is the engine of intelligence, powering every query, every agent and every application,” said Jensen Huang. “With NVIDIA Dynamo, we’ve created the first-ever ‘operating system’ for AI factories. The rapid adoption across our ecosystem shows this next wave of agentic AI is here, and NVIDIA is powering it at global scale.”

By improving throughput and reducing token costs, orchestration layers such as Dynamo reshape the economics of deploying agents at scale. Efficiency becomes a key differentiator, as enterprises move from experimentation to continuous operation of AI-driven systems.

Enterprise software enters an agentic phase

NVIDIA’s “full-stack” approach reflects a broader transformation in enterprise software architecture. Systems are evolving from static applications into dynamic environments where agents operate across layers, from data access to task execution.

This change introduces new complexities, particularly around integration, governance and system design. Enterprises must manage not only the performance of individual models but also the coordination of multiple agents operating within shared environments.

At the same time, the growing ecosystem of partners and integrations suggests that agentic systems are moving into production across industries. The focus is shifting from isolated pilots to scalable deployments embedded within core business processes.

The implications extend beyond productivity gains, redefining how software is structured and how work is performed. As agents become more capable and autonomous, enterprise software begins to resemble a coordinated system of interacting processes rather than a collection of tools.