We take one AI agent apart, layer by layer. Start with a plain agent that calls a single tool, then watch the architecture grow around it. I show each part you have to add before it survives production: tool boundaries, retries, timeouts, memory, and a human-in-the-loop checkpoint. For every layer, I explain the decision behind it and the failure it prevents. This is a guided build, not a framework tour, so the structure transfers to any stack. You leave with a clear mental model of how a real agent is wired end-to-end.