A chatbot answers you. An agent does things. The difference is not intelligence, it is permission and machinery: an agent is a model given tools, a goal, and the standing to take steps on its own. It reads the situation, decides what to do, acts, checks its work, and repeats, looping until the job is done or something needs a human.
That loop is the whole definition, and it is worth keeping in view because the word gets stretched over everything right now. A chat window that drafts text and waits for you is not an agent, however good the answers. A system that receives an email, extracts what matters, updates the record, drafts the response, and flags the odd case for review is one, because it took the steps itself.
For an operator, agents are where AI stops being a writing tool and starts being labor, which is also exactly why the discipline has to rise. A model that only talks can embarrass you. A model that acts can spend money, send email, and change records, so every real agent ships inside rules: which tools it may touch, what it may spend, where it must stop and ask. The capability question and the control question arrive together.
A concrete example. Invoices arrive in a shared inbox all day. An agent reads each one, pulls the details, checks them against the purchase order, files the record, and routes anything that does not match to a person. Each step is small and checkable. Chained, they hand back real hours on a task nobody wanted, which is the honest promise of agents: not magic, just work, done in a loop, with rules around it.
“Is this a chat feature or an agent, meaning does it actually take actions?”
// this page is the full text. the packaged PDF, all fifteen terms, comes with the free library.