Building AI Agents on Microsoft Foundry

What is Microsoft Foundry?

Microsoft Foundry (ai.azure.com) is Microsoft's studio for building AI applications. From a single place you deploy models, create agents, connect your own data, add tools, and ship to production - no separate infrastructure to wire up.

The key choice is how you create the agent. Foundry offers two paths: Build an agent (the Prompt Agent) and Code an agent (the Hosted Agent). This post covers both - and how to set each one up.

Two Ways to Build an Agent

Prompt Agent - you configure everything in the portal: instructions, tools, and knowledge.

Hosted Agent - your own Python code runs as a container inside Foundry.

Which one should you pick? Here is the short version:

Creating Your Foundry Project

Before building any agent, you create a project - the workspace that holds your agents, models, tools, and knowledge. Give it a name and, under advanced options, choose the region. We picked East US 2, since hosted agents are region-limited and not available everywhere.

Once the project is created, its home screen gives you the two paths - Build an agent and Code an agent - along with the project endpoint and model selection.

Setup 1: Build an Agent (Prompt Agent)

You can configure everything through the portal.

1. Open ai.azure.com and select (or create) your project.
2. Go to Agents -> New agent -> Build an agent.
3. Write the Instructions - the agent's behavior and rules.
4. Add Tools and Knowledge (web search, file upload, etc.) as needed.
5. Test it in the Playground.
6. Connect from your app via the project endpoint and Responses API.

Setup 2: Code an Agent (Hosted Agent)

Here you write code and deploy it as a container. Foundry handles the hosting; you use the Azure Developer CLI (azd).

1. Install the Azure Developer CLI (azd).
2. In Agents -> New agent -> Code an agent, copy the commands shown.
3. Run azd ai agent init to download a sample and scaffold the project (azure.yaml, the infra files, the Dockerfile, and the agent code).
4. Customize the code - add your own tools, instructions, and RAG.
5. Run azd up to build the container and deploy it to Foundry, provisioning any missing resources automatically.
6. The agent appears in the Agents list with type hosted, next to any prompt agents.

Example: A Sample Hosted Agent

Here is a sample hosted agent that combines RAG, custom tools, and memory - running on the gpt-5.4-mini model. A single user message can trigger several tools at once.

The trace shows exactly what ran: file search over the MDP document, exchange_rate, and get_weather each appear as a separate step before the model writes the final answer.

Evaluating the Agent

Foundry has a built-in Evaluations feature to test the agent automatically - quality, tool use, and safety - with no separate test harness. You pick the target agent, the test data (synthetic, an existing dataset, or real traces), and the evaluators; it then runs and reports a pass rate per metric.

Monitoring and Cost on Azure

Every resource lives in an Azure resource group, the Monitor dashboard reports agent runs, token usage, and estimated cost out of the box.

Azure Cost Management then breaks the spend down by service and resource - set a budget on the resource group to catch surprises early.

Conclusion

Microsoft Foundry makes building AI agents approachable from both ends. If you just need a quick assistant, a Prompt Agent gives you knowledge, memory, and tools in a few clicks. When you outgrow that and need your own code, custom tools, or tailored RAG, a Hosted Agent gives you full control on the same platform - with hosting, identity, evaluations, and cost tracking handled for you.