Entra Agent ID permission management isn't a UI problem

The confusion around why the API Permissions page seems to have vanished in Entra when configuring agent identities is driving people crazy, but it's not a bug or a missing feature; it's a fundamental shift in how we have to approach identity lifecycle management for non-human actors. Most people are looking at the dashboard expecting that old-school enterprise app workflow where you map a static set of permissions at provisioning and call it a day, but that's exactly how you end up with a security nightmare or a non-functional agent. If you follow the traditional path of granting broad, standing permissions to ensure the agent doesn't break, you're essentially violating the principle of least privilege from the jump, and if you're too stingy, the agent hits a permission wall the second the context shifts, which just leads to a flood of manual IT tickets. I've seen teams try to treat an AI agent like a standard service principal, but an agent is inherently dynamic—it moves across different data surfaces and tools based on the prompt context—so a static permission model is just too brittle to survive. Instead of hunting for a missing button in the portal, we had to rethink the governance architecture entirely. We moved away from the idea of "assigning permissions" to "managing capabilities" by using access packages as a sort of permission supply chain. In our implementation, the identity engineers define these packages as pre-governed capabilities, and the security teams can actually bake approval workflows into the process itself. This means the agent doesn't just sit there with a massive, permanent footprint; it requests what it needs for a specific task context through an automated path. It's a way to bridge that tension between the AI platform teams who want to move fast and the security teams who are rightfully terrified of an agent running wild. It turns the whole process from a one-time setup into a controlled, scalable lifecycle that actually accounts for how these models function under the hood.

https://entra.microsoft.com