Red Hat, NVIDIA ship OpenShell Reference Architecture for agents

The announcement builds on an existing product line, Red Hat AI Factory with NVIDIA, and extends it toward agentic AI rather than general model serving. The pivot matters because agents act on their own, so the two companies are pairing agent security features with optimized model, deployment and architecture work plus hardware support.

The OpenShell Reference Architecture is the most concrete artifact. By incorporating knowledge from the open-source OpenShell project, it offers a sandboxed runtime environment where guardrails can be standardized and operations made visible, which the companies frame as ensuring safety at execution time. Alongside it, a runtime combining vLLM with NVIDIA microservices is fully compatible with NVIDIA Nemotron 3.5 Super VL, which the companies say substantially cuts processing latency while maintaining accuracy.

The hardware side leans on NVIDIA's roadmap. Red Hat declared Day 0 support for the next-generation Vera Rubin architecture, and began supporting the agent-focused NVIDIA Vera CPU on Red Hat Enterprise Linux for NVIDIA. Distribution runs through Cisco, Dell Technologies and Lenovo, with the capabilities slated to be folded into future Red Hat AI platforms as standard features.