Overview
- Speaking on the Possible Podcast this weekend, Satya Nadella said AI agents need identities, sandboxes, permissions and governance and that he runs roughly 100 coding agents which create a high cognitive load to manage.
- Microsoft is packaging those controls as Agent 365 and plans to use Entra for identity and access and Purview for data labeling and tracing to contain and observe agent activity.
- Nadella argued that security, containment, manageability and observability are prerequisites for trusting agents that can write code, access data and act on users’ behalf.
- Reports note real operational risks from agent errors, such as mistaken deletions or misinterpreted instructions, which bolster calls for HR-like oversight, clear policies and audit trails for agent behavior.
- The comments follow Microsoft’s Build 2026 announcements and keep most agent features in private or limited previews, leaving wider availability, pricing and human-workflow integration unresolved.