The agent race is shifting from making agents smarter to fencing in and monitoring the environments they run in.
Over Sept 27–28, NVIDIA published its Open Agent Safety Platform and OpenShell runtime controls, Docker announced agent sandboxes that move between local and cloud, and Microsoft opened a preview of Copilot Managed Runtime, which runs code inside the same Microsoft 365 tenant. H Company's Holo4, a generalist computer-use agent, also landed in the same window.
Sandboxing, continuous monitoring and centralized admin control arrived from different camps at once. It suggests that as agents touch real work, whether they can run safely is becoming a condition of adoption, ahead of raw capability. Most items are announcements or previews, though, so real-world uptake is not yet confirmed.
Whether these platforms reach GA and interoperate, and whether computer-use agents like Holo4 start running inside these control layers in actual deployments.