A central agent policy decision works only when the host enforces it
Source: Microsoft
Microsoft's Agent Controls public preview lets teams define shared policies using the Agent Control Specification and evaluate proposed agent actions at hooks. The service can return allow, deny, or transform decisions with receipts. Microsoft's documentation says the host must supply trustworthy context and enforce each decision; the service does not sandbox an agent or govern paths that bypass the hooks.
Why this matters: Pilot one real tool call and one denied action. Check that the host actually stops the denied call, records the decision with enough context for an investigation, and keeps authentication and permissions in trusted systems. Then test an uninstrumented path. A central policy console cannot compensate for actions it never sees.
Read Microsoft's Agent Controls announcement