One sentence
A chatbot turns a prompt into text. An agent turns a goal into a sequence of tool calls and keeps going until the job is done or the budget ends.
Side by side
| Chatbot | Agent | |
|---|---|---|
| Output | Text | Actions + text |
| State | Single turn | Loop with memory |
| Tools | None | Shell, files, web, apps |
| Risk | Low | Higher — it acts |
The loop is the whole point
An agent plans, calls a tool, checks the result, and corrects course. That loop is why it can do work — and why it needs a budget and a sandbox.
When a chatbot is enough
For Q&A, summarising a doc, or drafting, a chatbot is simpler and cheaper. Reach for an agent when the task requires doing something: running commands, moving files, browsing, automating.
The middle ground
Most real work is a mix. That is why the agent’s controls — allowlist, budget, audit — matter more than the model underneath.