What separates it from a chatbot
A chatbot produces text and stops. An agent produces an action, observes what came back, and decides what to do next — a loop usually described as reason, act, observe. The model is not answering a question; it is working a task until a stopping condition is met.
That difference is what makes the category commercially interesting and what makes it a security concern. A system that can only produce text has a worst case of unhelpful text. A system holding tools has a worst case bounded by whatever those tools can reach.
The four components
A model that plans. It decomposes a request into steps and selects the next action, rather than emitting a single response.
Tools. Functions the model may invoke — query a warehouse, read a document, update a record. The tool roster defines the system's capability and, precisely, its blast radius.
Memory. State carried across steps and across turns, so a follow-up question does not restart from nothing.
Controls. The part most descriptions omit. Something has to decide which tools this user may reach, validate each proposed call before it runs, and stop the loop when it should stop. Without controls the other three components are a demo.
The question that separates production from prototype
Ask what happens when the agent decides to do something wrong. If the answer depends on the model choosing correctly, it is a prototype. If a gate refuses the action regardless of what the model decided, it is a system.
Where enterprises actually use it
The deployments that reach production share a shape: a bounded task, a defined set of systems, and a human approving anything that writes. Generating verification tests from a specification. Answering an analytical question by exploring a warehouse. Producing a first-draft dashboard from a request. Qualifying an account and proposing a CRM update.
What they have in common is that the mechanical portion of the work is large and the judgment portion is retained by a person. Deployments that attempt to automate the judgment are the ones still described as pilots eighteen months later.