Microsoft releases Foundry extension 1.0 to test and deploy AI agents

Microsoft is bringing agent setup and local testing into its Azure development workflow—not moving the whole process offline. On November 5, it released version 1.0 of the Microsoft Foundry extension for Azure Developer CLI. Developers can inspect an agent locally and deploy it to managed hosting, but model calls still run through Azure.

Azure Developer CLI, shortened to azd, creates Azure resources and deploys application code. The Foundry extension adds agent setup and local testing to that tool. In this workflow, a hosted agent runs the developer’s own code on managed compute. Developers choose the framework and the logic that coordinates the agent’s work; Foundry manages the runtime.

The connecting piece is azure.yaml, a configuration file stored alongside application code. It describes the agent and the Azure resources it needs, including model deployments, source location, runtime and compute allocation. Agent and deployment changes can be reviewed in the same pull request as code changes.

Teammates and automated build-and-deployment pipelines reuse that definition. Each azd environment supplies its own subscription, region and resource settings. Developers can begin without writing infrastructure files, then generate editable Bicep or Terraform files if they need to customize how resources are created.

Microsoft’s walkthrough requires Azure Developer CLI 1.37.0 or later, a current uv installation, an Azure subscription, Foundry permissions and model quota in the chosen region. Its sample uses Python 3.13 or later. After signing in, developers install the Foundry bundle with two commands.

azd auth login
azd ext install microsoft.foundry

Project initialization can use a new or existing Foundry project and reuse an existing model deployment. The provisioning step creates the declared resources, including missing model deployments in a reused project. Available deployment types and quota vary by subscription and region, so setup still involves choosing cloud resources rather than simply running a local sample.

Running azd ai agent run installs dependencies and opens Agent Inspector. Developers can inspect requests and responses while changing prompts, tools or code. When ready, azd deploy uploads the source to Foundry for a remote build, with no local container build required. The output includes playground and endpoint links.

The command set also includes a deployment preview for reviewing hosted-agent configuration changes, log streaming and a diagnostic command for setup or permission problems. Microsoft’s Foundry Skill gives coding assistants, including GitHub Copilot and Claude Code, workflows for project setup and local testing.

Quality testing uses a separately installed Evaluations extension. It can generate test data and a grading rubric, which developers review before configuring an evaluation and confirming the model that grades results. Test assets and configuration stay with the code for reuse in automated checks. An optimization command can use that evaluation as a baseline; Microsoft instructs developers to review candidates before explicitly deploying one.

For automated deployment, azd pipeline config sets up authentication, repository settings and a workflow to review and commit. Teams must grant the pipeline identity the required Azure permissions, keep runtime credentials in pipeline secrets and use a separate environment for each stage. Evaluation runs also need access to their referenced datasets and evaluators.

Cleanup has an important distinction. For a newly created demo project, azd down deletes the resource group and everything inside it. For a reused project, it leaves the project and its resources—including the hosted agent—in place. Unwanted resources then need separate removal.