What is an AI agent?

The core mechanism is a loop: the model reads the goal and the current state, decides the next action, calls a tool (an API, a database query, a file write), reads the result, and decides again. That loop can run for many steps without a person clicking through each one. What makes it an agent rather than a script is the decision-making in the middle: the model is choosing which tool to call and in what order based on what it just observed, not following a hardcoded sequence. This is why an AI agent is a different thing from a chatbot. A chatbot answers a message and stops. It has no persistent goal beyond the current turn and no ability to take an action outside the conversation window. An agent keeps working toward an objective across turns and can act on the world (send an email, update a record, book a meeting) as part of doing so. It is also different from a fixed workflow automation (think a Zapier-style trigger-then-action chain). A workflow runs the same steps in the same order every time, with no judgment call in between. An agent can hit an unexpected result, decide the original plan will not work, and route around it: try a different tool, ask for a missing piece of information, or break the task into new sub-steps that were not scripted in advance.