Your tenant already has agents in it. Some came from Copilot Studio, some from Azure AI Foundry, some from code nobody reviewed. Each one authenticates, holds permissions and touches data, and nobody can name most of them. So who is your agent, really?

This session answers that by following three agents through Microsoft Agent 365 and the Microsoft Entra Agent ID model. Reconciler is an autonomous Foundry agent that reads invoices from Storage and posts to an ERP overnight, no user in sight. Inbox assistant lives in Teams and summarizes your mail, acting on your behalf and never beyond it. Deskmate has its own mailbox and answers when you @mention it. Same identity machinery underneath, three very different tokens.

Live demos carry the session: we discover the agents already in a tenant, run Reconciler end to end from managed identity to blueprint to token to data, watch Inbox assistant act for a signed-in user under Conditional Access, and finish with the kill switch: one shared identity, every agent exposed, one switch to stop them all.

You leave able to inventory your agents, design blueprints per trust boundary, put permissions on the agent identity instead of the blueprint, and explain why publishing is a security boundary. With EU AI Act obligations landing in 2026, this is not a someday problem.

Level 300
~ 45 mins, depending on demos.

After this session, participants can:
Inventory the agents already in their tenant and treat each one as a non-human identity.
Design Entra Agent ID blueprints per trust boundary and scope agent permissions before deploying.
Tell on-behalf-of from own-identity access, and budget agents as identities, not seats.