What Microsoft announced
This October 19 catch-up is for AI geeks and experienced technicians. Microsoft describes Dynamic Permissions tied to an agent’s task and time limit, plus continuously updated Dynamic Agent Risk and Dynamic Conditional Access. Its examples include allowing reads while blocking writes as risk rises.
The announcement also introduces Playbooks and Guardian Agents in Project Perception. People set policy and approval boundaries; the agents help coordinate containment. These are Microsoft’s descriptions, not results from a Tinker & Tokens test.
Finding an agent is not the same as protecting it
Microsoft’s linked support matrix separates local-agent discovery from runtime protection. Windows runtime protection is marked preview; network inspection does not cover certificate pinning or HTTP/3. A discovery listing does not establish that protection is enabled or that every action is inspected.
The Entra Agent ID overview also separates creating agent identities from extending security features to agents, which requires Microsoft Agent 365. Check the relevant license, supported application, inspection method, and configured mode before relying on a control. The announcement does not establish universal availability for every new feature.
Our take
R2’s editorial analysis: start with the specific job. An assistant sorting draft invoices may need to read a folder; that alone is no reason to give it authority to pay, send, or approve them. Give the task a narrow scope and an expiry, with consequential actions checked outside the model.
A second AI watching the first can help, but it also needs enforceable limits. Useful self-check: can you show which action a control actually blocked, and who can restore access? A product announcement is a reason to evaluate that question, not evidence that your workflow is already protected.