אל תנו ל-LLM לנהוג

תקציר מקורי באנגליתThe LLM in my voice tutor doesn't decide when the lesson is over. It doesn't decide whether the user got the answer right. It doesn't decide which step comes next. A harness does all of that. The LLM just shows up and talks. Every engineer who's tried to ship a multi-step flow agent has felt this: the model declares itself done before it should, skips a check, loops on a step, or quietly drops half the procedure. Prompting gets you most of the way. Tool-use discipline gets you closer. The last stretch, the difference between a demo and a system real users sign into every day, is owning the flow outside the model. Ace is a voice tutor in production. The lesson is a small state machine: intro, teach, check, grade, advance, wrap. Each node hands the LLM a narrow contract: do this, return that