What an AI agent is (and what it is not)

A chatbot answers a question and stops. An agent is given a job, some tools, and permission to loop until the job is done or it hits a limit.

The useful parts

• Tools. Fetch a page, query a database, run a command the human allowed. Without tools, it is still a chatbot with extra steps.
• Loop. Observe, act, check. The loop is the job. Cap the number of steps. An unbounded loop is not autonomy; it is a hang.
• Memory. What it may keep (the current ticket, the last filing) and what it must not (secrets, other customers data).
• Limits. What it must refuse. If you skip this, the agent will fill the gap with guesses.

What the label does not mean

Calling it an agent does not mean it knows your business, has authority, or should be left unsupervised. It means you described a job with tools and a stop condition.

If it cannot use the tools, tell it to stop. Inventing a filing, a price, or a CVE proof is a failure, not a feature.