A useful mental model for AI agents The agent should not be the workflow. A lot of agent systems still look like - model → decide → tool → decide → tool → repeat It works, but giving the model control over the entire workflow creates unnecessary uncertainty. Some parts are better handled deterministically. For example - The model can decide which customer request it's dealing with. Your application can decide the authentication rules. The model can suggest an action. Your code can validate whether that action is allowed. The model can choose the next step. Your workflow engine can enforce ordering, retries, timeouts and state transitions. This creates a much better separation. LLM = reasoning Application = control Tools = execution The more important the operation, the less you should rely on the model to enforce the rules. Good agent architecture isn't about giving the model maximum control. It's about giving it exactly enough control to be useful.