A chatbot is a conversational interface; its internal behavior may be simple or complex. Retrieval-augmented generation finds relevant evidence and supplies it to a model. A fixed workflow follows a sequence chosen by application code. An agent can choose some next steps based on observations. A single product can use all of these patterns.
Consider answering a policy question. A fixed workflow searches once and drafts from the results. An agent might decide that the evidence is insufficient and search again. Both can have a chat interface. The visible interface does not tell you how much autonomy the system has.
Do not add an agent merely because a task uses a language model. If the required sequence is known and reliable, ordinary code may be easier to test and operate. Adaptive behavior should solve an observed problem and remain bounded.
Exercise: draw the steps for a document question-answering feature and mark which steps are fixed and which require a decision based on new information.
Check: you can defend the simplest design that meets the task, including a design with no agent loop.