Treating an agent deployment like "a web app that happens to call an LLM" will backfire in production.