A chatbot answers, an agent acts
A chatbot turns a prompt into text. An agent turns a goal into a series of steps: it reads a task, decides what to do, calls tools like a shell or a browser, checks the result, and keeps going until the job is done.
That loop is the whole difference. It is also why agents are more useful — and why a mistake is more expensive, because one wrong step can delete a file or run up a bill.
The loop, in plain terms
- Receive — a chat message, a scheduled job, or a goal.
- Plan — pick a strategy and a step budget.
- Act — call tools: shell, files, web, apps, browser.
- Reflect — compare the result against the plan.
- Verify — check the answer before you see it.
- Record — write the trace to the audit trail.
Where the API key fits
Every step the agent takes is a model call, and every model call needs credentials. If the key is embedded in the agent, then a bad tool call can expose it. If the key sits behind a proxy, the agent never has the secret to leak.
Autonomy needs guardrails
An agent that runs on its own needs boundaries: an allowlist of tools, a token budget, and a log of everything it did. Without those, "autonomous" just means "unattended and hard to audit".
Agents vs assistants vs workflows
| Term | What it does | Runs |
|---|---|---|
| Assistant | Answers questions from context | On demand |
| Workflow | Fixed steps, no judgment | On a trigger |
| Agent | Plans and adapts in a loop | Until the goal or budget ends |