Learn · 3 of 7
Generative, agentic, and the gap between them
The difference between a model that writes and a model that acts, and why it matters for your bill.
Both words get used loosely enough to be worth pinning down.
Generative describes what the model produces: new text, rather than a label or a score. Ask it to write an email and you are using it generatively. That is most of what people do.
Agentic describes the arrangement around it: the model is given tools (Letting a model call functions you define — search, read a file, send a request — and read back what they return.) and a goal, and loops — deciding what to do, doing it, reading the result, deciding again — without anyone approving each step.
The difference is autonomy, not intelligence
It is the same model in both cases. What changes is who decides the next step.
In a chat, you do. You read the answer, think, and type again. In an agent loop (Plan, call a tool, read the result, decide again. The cycle an agent repeats until it finishes or is stopped.), the model does, and keeps doing it until it finishes or something stops it. That is where the usefulness comes from — it will work through twenty steps while you do something else — and all of the risk, because an agent that has misunderstood the goal pursues the wrong thing energetically.
It changes the economics completely
This is the part that catches people out, and the diagram above shows it.
Each tool result is appended to the context, so the context grows on every pass. Turn one might send five thousand tokens; turn twenty might send eighty thousand, because it is carrying nineteen results. You pay for all of it, every turn. The cost of a loop is closer to quadratic in its length than linear.
And because the context changes each pass, the prompt cache (Paying a reduced rate for a prefix the vendor has already processed, instead of full price for sending it again.) is invalidated each pass — which is why cache writes tend to dominate an agentic bill rather than output.
Which to reach for
If a person is going to read every answer anyway, a chat is cheaper and more controllable. If the task is twenty mechanical steps where each one depends on the last, an agent earns its cost. The mistake is using an agent for something a single well-aimed prompt would have done — paying loop prices for a one-shot job.