An agent needs more than a model: it needs a clear task, limited data sources, permitted actions and a route to human approval. Microsoft Foundry provides projects, models, agent tools and evaluation. Even so, the first design should start with a single workflow you can check.
A small, useful pilot
Take answering internal IT questions as an example. Put together ten real questions, mark the correct sources and decide when the agent hands over to the help desk. Then connect only the approved knowledge base. An agent that also changes tickets and sends emails makes this test unnecessarily hard to check.
Four boundaries before go-live
Check the model deployment, where the agent state is stored, where external tools run and the permissions on connected data sources. According to Microsoft's agent data privacy documentation, where model and tool calls are processed depends on how they are configured. A Swiss Azure resource alone is therefore no assurance that every data flow stays in Switzerland.
Use the current SDK documentation
For new development, use the current SDK and endpoint documentation. Older code samples that connect through project connection strings do not match current Foundry projects. Only move an agent snippet into a production application once you have tested it with your SDK version.
Deciding after the pilot
Compare correct answers with cited sources, impermissible answers, response time and cost. Give further teams access only once these figures are acceptable.

