The next phase of artificial intelligence will not be defined by better models, but by systems that can act. That was the underlying message from NVIDIA at its GTC conference, where the company set out a vision of enterprise software evolving into networks of autonomous agents capable of reasoning, planning and executing tasks with minimal human input.
At the centre of this transition is the introduction of the NVIDIA Agent Toolkit, a collection of open models, frameworks and runtime environments designed to enable enterprises to build and deploy AI agents that can reason, plan and execute tasks independently. The toolkit includes OpenShell, an open source runtime that enforces policy-based controls, allowing organisations to govern how agents access data, interact with systems and operate within defined security boundaries.
This reflects a broader change in how AI is being integrated into enterprise environments. Rather than acting as tools that respond to prompts, agents are being positioned as systems that can determine how to complete tasks, selecting data sources, orchestrating workflows and interacting with other software autonomously.
From tools to systems
The emergence of agentic AI introduces new technical and operational challenges. Systems must not only generate accurate outputs, but also manage long-running processes, maintain context across multiple steps and operate securely within enterprise environments. NVIDIA’s approach focuses on combining open models with orchestration frameworks to create systems that can scale across these requirements.
The expansion of the Nemotron model family reflects this direction, with multimodal models designed to process language, vision and audio inputs while supporting complex reasoning and real-time interaction. These models are being adopted across a range of applications, from coding assistants to enterprise workflow automation, as organisations look to embed AI more deeply into their operations.
At the same time, the formation of the Nemotron Coalition signals an effort to accelerate the development of open foundation models through collaboration between AI labs and developers. By pooling data, compute and expertise, the initiative aims to create shared models that can be adapted to specific industries and regional requirements, reflecting growing demand for more specialised and controllable AI systems.
The operating system for agents
Alongside model development, NVIDIA is also addressing the infrastructure required to run autonomous systems continuously. The NemoClaw stack, designed for the OpenClaw agent platform, introduces a simplified deployment model that allows developers to install and run AI agents with built-in privacy and security controls. These agents can operate across both local and cloud environments, combining open models with external services to complete tasks while maintaining policy-based safeguards.
This approach points to the emergence of a new layer in enterprise computing, where operating systems are no longer limited to managing hardware and applications, but are extended to govern the behaviour of autonomous agents. As these systems become more capable, the challenge shifts from building individual models to coordinating large numbers of agents working across complex workflows.
The implications are significant. If agents can plan, execute and refine tasks independently, the role of software begins to change. Applications become collections of capabilities that agents can draw upon, rather than fixed interfaces that users interact with directly. In this model, productivity gains are driven not only by faster computation, but by the ability of systems to act without constant human intervention.
The announcements at GTC suggest that this shift is already underway. With enterprises beginning to adopt agentic systems and the supporting infrastructure taking shape, the question is no longer whether AI will move beyond generation into action, but how quickly organisations can adapt to a world where software is no longer static, but continuously evolving and increasingly autonomous.
