The Agentic Application Architecture

Inside a real application an agent is one component among many. There is the model; the prompts that shape it; the tools it may call; the state of the current task; memory across tasks; retrieval; orchestration that decides which step runs next; ordinary application logic; the APIs it talks to; and the interface a person sees. Most of these existed long before language models did.

Drawing this diagram for your own application is the fastest way to see where the uncertainty actually lives. Usually it is in two or three boxes. The rest is software you already know how to build and test.