Quick answer: A chatbot can only respond within a single exchange, with no memory of prior context and no ability to take multi-step action. An AI agent adds three capabilities a basic chatbot lacks: memory across a session, the ability to plan multi-step tasks, and the ability to actually execute actions rather than just respond.
Early chatbots could only respond within a single exchange. Getting from there to systems that can complete multi-step tasks required three specific technical capabilities to be added, not just bigger models.
The first shift was models that could reference earlier parts of a conversation instead of treating every message as a fresh start - the basic prerequisite for anything beyond a one-shot question and answer.
The next capability was letting a model call an external tool - a search, a calculator, a code interpreter - and use the result in its response, rather than relying only on what it learned during training.
The most recent shift is models that can break a goal into steps, execute them, check whether a step worked, and adjust if it didn't - the difference between "answer this question" and "complete this task."
Each of these capabilities can be used independently. A workflow doesn't need full autonomous planning to be useful - a tool-use-only setup that reliably automates one specific step is often the more practical starting point.