Red Hat & NVIDIA expand AI Factory with agent controls

Red Hat has expanded its AI Factory collaboration with NVIDIA, focusing on security controls for AI agents, model inference design and support for NVIDIA hardware.

The update includes work on the OpenShell Reference Architecture, a joint inference design using vLLM and llm-d, and launch support for NVIDIA Vera Rubin infrastructure and NVIDIA Vera CPUs.

Agent controls

A key part of the announcement is agent security. Red Hat is working with NVIDIA on OpenShell, an open-source agent sandbox runtime designed to set boundaries on what agents can do through controls enforced outside the agent itself.

The aim is to add runtime safety, visibility and standard guardrails for autonomous AI agents used in business systems. Red Hat and NVIDIA said they are leading contributors to the OpenShell project, and that the work will be supported through Red Hat AI Factory with NVIDIA.

The focus reflects a broader shift in the AI market as companies move from chat-based tools to software agents that can take action inside IT environments. These systems place greater weight on control mechanisms because they can reason, call tools and execute tasks with less direct human intervention.

Joe Fernandes, Vice President and General Manager of the AI Business Unit at Red Hat, said: "Enhancing the security of agentic AI will fuel the next major evolution in enterprise computing, but scaling autonomous agents requires trust, safety and seamless infrastructure integration. By deepening our collaboration with NVIDIA across open-source communities like OpenShell, delivering optimised inference architectures and driving day-zero hardware enablement, we are giving IT decision-makers the trusted foundation they need to confidently deploy AI agents at scale."

NVIDIA described the issue in similar terms, emphasising the separation of policy enforcement from the agent itself.

"An agent shouldn't be enforcing its own boundaries - those controls belong outside the agent, in the runtime and the infrastructure beneath it. NVIDIA's expanded work with Red Hat combines a secure agent runtime, governed inference and deep hardware support to give enterprises a trusted, high-performance blueprint for agentic AI across the open hybrid cloud," said Fanelli.

Inference design

Red Hat and NVIDIA also outlined a reference design for model inference built on Red Hat AI Factory with NVIDIA. It combines Model-as-a-Service access control, an AI Gateway, vLLM and llm-d, alongside NVIDIA Dynamo, NVIDIA NIM microservices and GPU orchestration.

The companies said the setup is based on the NVIDIA Inference Hub reference design and is intended to let organisations discover models through self-service while applying access policies, rate limits and subscription-based entitlement tracking.

The design also separates model access governance from execution. This is intended to give IT teams more direct control over who can use models, how often they can be called and how usage is tracked across hybrid cloud environments.

Red Hat said the latest NVIDIA Nemotron models, including NVIDIA Nemotron 3.5 Super VL, are supported in vLLM and in Red Hat AI Factory with NVIDIA. It added that quantisation techniques through its LLM compressor have been demonstrated with FP8 and DSpark speculative decoding checkpoints for Nemotron 3.5 Super VL.

Hardware support

The third element of the expansion is hardware support. Red Hat said it will provide day-zero enablement for NVIDIA Vera Rubin infrastructure, allowing its software stack to support the platform when it becomes available.

Red Hat Enterprise Linux for NVIDIA also now supports NVIDIA Vera CPUs. The product is a version of Red Hat Enterprise Linux tailored for NVIDIA systems.

The hardware tie-up points to the growing importance of infrastructure tuning as companies test and deploy AI agents at scale. Agentic workloads can place continuous demands on compute resources because they involve repeated reasoning steps, tool use and sandboxed execution rather than one-off prompts.

Partner access

Red Hat said the joint use cases, reference designs and agentic AI features are available through its partner network, including Cisco, Dell Technologies and Lenovo.

It also said it is working with NVIDIA and the wider community to integrate the OpenShell Reference Architecture into Red Hat AI and Red Hat AI Factory with NVIDIA as a native platform feature.

For enterprise buyers, the announcement shows how major infrastructure suppliers are trying to shape the operational layer around AI agents, not just the models. The emphasis on guardrails, access governance and hardware support suggests vendors expect customers to focus as much on control and system design as on raw model performance.