Most teams today know how to build an agent. Fewer teams have a plan for what happens after go-live. An agent should be treated like any other application, which means someone has to manage its lifecycle, and the same is true for the large language model behind it. If you have an agent in production, you should know the retirement date of the model it runs on, what happens to the agent when that model changes, and how you will test and fix it before that date. In this one-hour session, Isha Kapoor, David Lorenzo, and I walk through a practical framework that answers these questions.

We start with the agent lifecycle in Copilot Studio, which covers controlled deployment, rollback, quarantine, and decommissioning. Then we move to identity, because agents can now have their own identity in Microsoft Entra. We show how to govern that identity with access packages, sponsors, attestation, and Lifecycle Workflows, how to keep connector and API scope under control, and how Agent 365 helps with managing the agent lifecycle. In the last part we cover the AI model lifecycle, which means how to decide when to upgrade a model, how to run and validate a model migration for Copilot Studio agents with evaluations, and how version policy works for model deployments in Microsoft Foundry.

The session is for architects, IT admins, and makers who already build agents and now need to govern and operate them. All references point to official Microsoft Learn documentation, and the supporting files are available on GitHub. Watch the full video below, and subscribe to the YouTube channel for more sessions like this.