Moving from a Copilot-style assistant to an AI agent changes the architecture dramatically. The system no longer just generates an answer; it must understand context, choose tools, retrieve enterprise data, make decisions, execute actions, and recover when something goes wrong. In this demo-driven session, we will build an enterprise agent step by step and examine the architecture behind reasoning, grounding, tool calling, memory, state, permissions, and human approval. We will deliberately introduce failure scenarios such as incorrect tool selection, insufficient context, repeated actions, excessive permissions, and low-confidence decisions, then show patterns for controlling them. The session will also demonstrate where deterministic workflows should remain deterministic rather than being replaced by AI. Attendees will leave with a practical architecture for moving from conversational assistants to useful enterprise agents while keeping execution observable, bounded, and safe.