Run secure agent loops in production with isolated sandboxes at scale | BRK448

Christian Flasshoff, Jeffrey Quesnelle, Devanshi Joshi and Jan Kalis examine the infrastructure needed to move agent workloads into production, with Azure Container Apps sandbox demonstrations and enterprise deployment examples.

Production agents need a platform, not just a model

The session frames production AI agents as a platform problem. Before organizations can trust agents to run workloads, they need secure isolation, identity, governance, observability and cost control.

Azure Container Apps provides the foundation discussed in the session, combining production-ready defaults with microVM isolation and governance. The presenters also explain how Copilot Studio and Foundry Agent Service leverage this architecture and how enterprises can implement it.

Isolation and enterprise controls

The session examines why sandboxed microVMs are becoming a foundation for agent workloads, then connects that infrastructure to practical deployment concerns:

Sandbox demonstrations use the Azure portal, CLI and Visual Studio Code.

Deployment examples: Hermes Agent and Cognite

The Hermes Agent segment covers its development as a production AI agent platform, self-improvement loops, and the governance and security challenges those loops introduce. It also examines how Hermes Agent uses Azure Container App Sandboxes, including its architecture, enterprise controls and lessons from large-scale deployments.

The Cognite segment focuses on industrial AI use cases powered by secure sandboxed agents, alongside insights from production deployments.

Session chapters

Resource

This intermediate-level Cloud & AI breakout session, BRK448, is part of Microsoft Ignite 2026 and is presented in English (US).