4.2.2. Creating and Testing a Single Agent
💡 First Principle: Creating an agent means defining its identity and behavior — a name, the model it uses, and its instructions (effectively a system prompt for the agent). The fastest path is the Foundry portal's agent playground: you give instructions in plain language and immediately chat with the agent to test it, no code required.
In the portal you create an agent, set instructions like "You are a helpful agent that answers questions about geography," and start chatting to verify it behaves. This is the no-code prompt agent path. The same agent can then be created or managed programmatically through the project client's agent APIs, but for the exam the key idea is: define instructions + model, then test by conversing with it before wiring it into an app.
⚠️ Exam Trap: An agent's instructions play the same role a system prompt does for a chat model — they set persistent behavior. Vague instructions produce an unreliable agent. "Be helpful" is weak; "Answer only geography questions; if asked anything else, politely decline" is testable and bounded.
Reflection Question: Why is testing an agent in the playground (by actually chatting with it) more revealing than just reading its instructions? What kinds of problems does conversing surface?
