Azure agentic apps: secure execution and durable workflows
Compute

Azure agentic apps: secure execution and durable workflows

Azure developers can combine Container Apps sandboxes, Functions hosted skills and durable workflows to build agent-enabled applications.

Key insights

  • Azure Container Apps Sandboxes provide hardware-isolated microVMs for running agent tools and code.
  • Functions hosted skills package agent instructions in Markdown and connect them to event triggers and bindings.
  • Durable Task Scheduler persists workflow state and can recover execution after a crash.
  • Azure Connector Namespace makes Logic Apps connectors available as a separate Azure resource.
  • The session describes sandboxes as generally available and hosted skills as available now.

Azure’s agentic application platform brings together hosting, isolated execution, integrations and durable orchestration so developers can build applications that use agents without treating every task as a chat interaction. It matters to Azure developers and platform teams that need to run agent tools, connect business systems or keep long-running workflows recoverable. Microsoft Events demonstrates how these pieces fit together in a Microsoft Ignite session.

What it does

The platform is not presented as a single new service. Instead, it combines Azure App Service, Azure Container Apps and Azure Functions with building blocks for agent execution, workflow state and integrations. The aim is to let teams add agent capabilities to applications on Azure rather than requiring a move to a separate hosting platform.

Each service has a different role in the design. App Service can remain part of an existing application, Container Apps provides a place for container-based workloads and isolated sandboxes, and Functions can host event-driven skills. Durable Task Scheduler addresses a different problem: keeping progress in workflows that run for a long time, pause or need to recover.

That distinction matters because an agent may call tools or run code, not just return text. The session’s design principles include giving each agent its own identity, isolating code execution, monitoring and controlling network egress outside the sandbox, and preventing sandboxed code from accessing secrets directly.

Who it is for and what you need

The components are aimed at developers building agent-enabled applications and platform teams providing secure execution for them. They are also relevant to teams extending existing event-driven applications with agent behavior, or connecting agents to business systems through reusable integrations.

The session describes Azure Container Apps Sandboxes as generally available. It describes hosted skills in Azure Functions as available now. Durable Task Scheduler is presented as a managed workflow engine supported across Azure Functions, Azure App Service, Azure Container Apps and Azure Kubernetes Service.

The approach is composable: the speakers describe using the capabilities with existing Azure services and Microsoft Foundry. The discussion also describes sandboxes as a foundation used by Microsoft Foundry-hosted agents, Copilot Studio, GitHub Copilot workspace, Azure SRE Agent and Azure Container Apps Express.

For teams evaluating the Azure compute services, the practical starting point is to identify the workload’s needs: does it need isolated code execution, event-based activation, durable state, or access to business-system connectors? The services address different parts of that application rather than replacing one another.

How it works

  1. Run risky or tool-generated code in a sandbox. Azure Container Apps Sandboxes use hardware-isolated microVMs. The presentation describes short-lived and long-running workloads with preserved context, as well as bursty execution without idle compute cost.
  2. Respond to events with a hosted skill. In Azure Functions, a skill is described as instructions written in Markdown. Events such as a queue message, timer, webhook, alert or email can start the skill. Functions triggers and bindings connect the work to other data or actions.
  3. Persist work that cannot finish in one run. Durable Task Scheduler stores workflow progress so an execution can recover after a crash or resume after infrastructure scales to zero. The session also describes long-running transactions and human-in-the-loop steps.
  4. Connect business systems. Azure Connector Namespace makes the Logic Apps connector ecosystem available across Azure as its own resource. The session says connectors can be given a user-assigned or system-assigned identity and used to compose applications.

There is also a connection between hosted skills and durable workflows. The presenters describe a capability that reads a skill’s Markdown and generates a coded Durable Task version. The stated goal is to turn non-deterministic instructions into a deterministic workflow graph and reduce token use on subsequent runs.

For operators, the scheduler includes telemetry and a dashboard that shows workflow progress. At the execution boundary, the design calls for monitoring network access outside the sandbox, where restrictions can be applied without trusting the code running inside it.

Limitations and alternatives

These building blocks address different requirements, so teams should choose based on the work an agent must perform. A hosted skill responds to events and can be added to an existing Functions app; it is not the same thing as a persistent agent service or a durable orchestration engine. A sandbox supplies an isolated place to execute code, while Durable Task Scheduler manages workflow progress.

Security also depends on how the components are used. The design described in the session avoids giving sandboxed code direct access to secrets and puts egress monitoring outside the sandbox. Teams should account for identity, permitted network access and the actions an agent is allowed to take when designing their own implementation.

The speakers emphasize that these capabilities can be composed with existing Azure services rather than requiring an application migration. The session does not provide a universal setup recipe or configuration values; implementation choices depend on the application’s hosting, integration and workflow needs.

Sources

Frequently asked questions

What are Azure Container Apps Sandboxes?

They are hardware-isolated microVMs for running agent code and tools. The session describes short-lived and long-running workloads with preserved context.

How do hosted skills work in Azure Functions?

A hosted skill uses Markdown instructions and can be triggered by events such as a queue message, timer, webhook, alert or email. Functions triggers and bindings connect it to inputs and actions.

What does Durable Task Scheduler do?

It persists workflow progress and can recover execution after a crash or resume after infrastructure scales to zero. It also supports long-running transactions and human-in-the-loop steps.