Attention on agents has moved from "can you build one" to "can you run one in production safely". What shipped this week was runtime defence and governance at scale.
NVIDIA released Sentry, which protects agents while they run; the OpenShell reference architecture, which governs agent workloads at scale; and Nemotron 3.5 Embed, Rerank and Parse, which handle retrieval quality. Alongside these it published a workflow for building an AI factory digital twin with agent assistance. None of these is a new generative model — they are the layer around the agent.
The means of building agents are already widespread; the wall comes after you ship. What the agent may not touch at runtime, whose credentials it acts under, and whether retrieval hands it wrong evidence — until those are settled, a working demo does not become a deployment. This week's releases map onto exactly those three. Note: how much they help in practice is unknown until adopters outside NVIDIA report back.
Whether deployments of Sentry and OpenShell surface from operators other than NVIDIA. Whether Nemotron 3.5's retrieval quality is enough to displace existing RAG stacks.