What is an AI agent (and when you actually need one)

A prompt, an automation, and an agent are three different things

A prompt is one question and one answer. You supply the input, the model produces text, you decide what happens next. Nothing runs without you.

An automation is a fixed pipeline. When this happens, do that, then that. It is reliable because it never deviates — and useless the moment the input does not match the shape you planned for.

An agent sits between them. You give it a goal instead of a sequence, hand it a small set of tools it may call, and let it decide the order. It loops: think, act, look at the result, decide whether it is done. That loop is the only real difference, and it is also the source of every problem agents have.

• Prompt — one step, you drive.
• Automation — many steps, fixed order, no judgment.
• Agent — many steps, chosen order, judgment at each step.

The four parts every agent has

Whatever you build it in, an agent is made of the same four pieces. If you can name all four for the thing you are building, you understand it well enough to ship it.

• Instructions — the job, the standards, and the limits. Written once, reused every run.
• Tools — the specific actions it may take: search, read a sheet, send a draft, call an API. Nothing else is possible.
• Memory — what it can see. The current task, and optionally the history of previous ones.
• A stopping rule — how it knows it is finished, and what happens when it is not.

The stopping rule is the part everyone skips

An agent without a done-condition either quits early or spins. I write the stopping rule before the instructions now, because it forces me to say what a good result actually looks like — and half the time, writing it down reveals the job did not need an agent at all.

A stopping rule should be checkable by something other than the model's opinion. 'Summary written' is not a stopping rule. 'A row exists in the sheet for every item in the source list, and each has a non-empty status' is.

I want to build an AI agent that does the following job: {describe the job in one paragraph}.

Before any building, do three things:
1. Write the done-condition — a check someone else could run, without judgment, to confirm the job finished correctly.
2. List the smallest possible set of tools the agent needs. Justify each one.
3. Tell me honestly whether this job needs an agent at all, or whether a single prompt or a fixed automation would be more reliable. If it does not need an agent, say so and describe the simpler version.

When you do not need an agent

Most of the jobs people describe to me are one-step jobs wearing a costume. If the work is the same every time and the input is the same shape every time, an agent adds cost, latency, and a new way to be wrong. Use the fixed version.

I reach for an agent when the number of steps depends on what it finds. Researching a topic where the good sources are unknown up front. Triaging a message where the right action depends on what the message says. Extracting data from documents that are all laid out differently. In those, the branching is the job.

• Same input shape, same steps every time → automation.
• One question, one answer → prompt.
• Steps depend on what it finds → agent.
• Wrong answer is expensive and irreversible → agent with an approval step, or no agent.

What an agent costs you that a prompt does not

Every loop is another model call, so cost scales with how confused it gets. A well-scoped agent finishes in three or four steps; a vague one wanders through twenty and produces something worse.

You also inherit a new failure mode: confident wrong action. A prompt that hallucinates hands you bad text and you notice. An agent that hallucinates calls a tool with bad arguments and the mistake is now in your sheet, your inbox, or your customer's inbox. That is why anything that sends, spends, or deletes gets a human approval step in everything I run.

The honest limit

Agents are good at the work between the interesting parts — gathering, sorting, drafting, checking, formatting. They are not good at deciding what matters, and they do not know your business. Everything I run produces a draft or a shortlist that I approve. That is not a lack of ambition; it is the configuration that has survived contact with real work.

If you want the builds rather than the concepts, the next page walks a complete agent end to end with no code, and the course covers six of them at three depths.