A chatbot answers. An agent acts. That difference sounds like marketing until you look at what it changes in the engineering, which is almost everything about how the system has to be tested and constrained.
What actually defines an agent
An agent can call tools, query data sources, decide between steps within defined boundaries, and retry when a step fails. It does not need a human prompting each individual action. That autonomy is the whole value and the whole risk.
Why evaluation stops being optional
A chatbot that answers badly wastes a user's time. An agent that acts badly writes to your database. Before building the execution layer we build a scored evaluation set from real tasks with expected outcomes, and every change is measured against it — otherwise there is no way to tell whether a model upgrade helped or quietly regressed.
The boring parts that make it work
Most of the engineering is not the model. It is idempotent tool calls, so a retry cannot double-charge someone. It is explicit permission scoping, so the agent cannot reach data the requesting user could not. It is hard ceilings on spend and step count. None of it is glamorous, and all of it is what separates a demo from something you can leave running.
Where a human still belongs
Decisions with legal, financial or safety consequences keep a person in the loop, and the interface shows the model's uncertainty rather than hiding it behind confident prose.