Build AI agents with Foundry, RAG and tools
Build an AI application from the model call through retrieval, tools and evaluation. The examples use Microsoft Foundry, Azure OpenAI and agent tooling such as MCP. Some articles explain architecture choices; others follow a portal walkthrough or include code. Choose the step that matches your project, then check identity, permissions and operation before extending it.
Locate the part of the agent you are building
Start with project and model deployment if you do not yet have a working model call. Move to retrieval when the answer must use an approved set of documents. Tool integration addresses actions or live data that a model cannot obtain from its instructions alone. Evaluation and tracing help you understand what happened after the first successful response. These are separate tasks, so choose a guide for the next missing capability instead of adding every feature at once.
Keep the first example bounded
Define one input, the information the agent may use and an observable result. Follow the article with a small source set and an identity whose access you can explain. For a tool call, distinguish reading information from changing a system. Record the expected response before running the example and inspect the returned evidence afterwards. A fluent answer is not proof that the agent selected the correct source, respected a permission boundary or completed an action.
Check operation before extending the prototype
Read the dependencies, model name and publication date alongside the example. Product availability and API versions may differ from the environment in a video. Use the linked documentation to confirm the prerequisites. When the example works, retain a trace of the call and test a missing source, an unavailable tool and a request outside the allowed scope. The related guides cover these checks in more detail and show where retrieval, identity, tools and evaluation meet.