4.2.1. What an Agent Is
💡 First Principle: An agent = a model (the reasoning engine) + instructions (its role and goal) + tools (things it can do). Given a request, it loops: reason about what's needed, optionally call a tool, observe the result, and continue until it can answer. That loop is what makes it "agentic."
Consider "book me a table for two tomorrow at 7." A plain model can only produce text about booking. An agent with a reservations tool can actually check availability and make the booking, then confirm. The agent uses the model to decide which tool to call and when, but the tools are what let it affect the world. Foundry's Agent Service hosts these agents and handles scaling, identity, and safety guardrails (including mitigation of prompt-injection attacks — the XPIA concern from Phase 2).
⚠️ Exam Trap: Not every AI feature needs an agent. If a task is "generate a summary" or "answer a question from given text," a plain model call is the right, cheaper choice. Reach for an agent when the task requires taking actions or calling tools to complete a goal.
Reflection Question: Classify each as "plain model call" or "agent": (a) summarize this email; (b) check inventory and place a reorder if stock is low. What distinguishes them?
