What Is an AI Agent? (And What It Isn’t)

The app built in the last lesson followed a fixed pattern: input in, one AI call, response out. This module is about what changes once that pattern breaks — once the model itself starts choosing what happens next. "Agent" has become the most overloaded word in AI to describe that shift. A chatbot with a personality gets called an agent. So does a scheduled script. So does a system that genuinely plans and acts on its own. These are not the same thing, and confusing them is how teams end up building something far more complicated than they needed.

Three different things people call "agents"

It helps to put them on a ladder, from simplest to most complex. Almost every real product sits on one of these three rungs:

• A single call. You send text, you get text back. Summarize this, classify that, answer this question. There is no loop, no memory beyond what you sent, no decisions. Most useful AI features are this — and that is a good thing.
• A workflow. Multiple AI calls chained together, but you wrote the sequence: first summarize, then extract action items, then draft an email. The path is fixed in your code. The AI fills in steps; it doesn't choose them.
• An agent. The model decides what to do next. You give it a goal and a set of tools, and it loops: think, act, observe the result, think again — until it decides the job is done. You did not write the sequence, because the sequence depends on what it finds along the way.

Where the line actually is

The dividing line is not intelligence, autonomy-as-a-vibe, or how impressive the demo looks. It is one concrete question: who decides the next step — your code, or the model?

If your code contains the sequence, it's a workflow, however sophisticated each step is. If the model picks the next action based on what the last action returned, it's an agent. That's the whole distinction. A three-line loop can be a genuine agent; a thousand-line pipeline with a dozen AI calls can still be a workflow.

Why the distinction matters

Because each rung up costs you something real. A single call is fast, cheap, and predictable — you can test it exhaustively. A workflow is slower and pricier but still deterministic: the same input follows the same path, so when it breaks you know where to look.

An agent gives that up. It may take three tool calls or thirty. Two identical requests can take different paths. Costs vary per run, latency is unpredictable, and debugging means reconstructing a decision chain rather than reading a stack trace. That trade is worth making — but only when the task genuinely can't be specified in advance.

The honest guidance is to use the simplest rung that solves your problem. Reach for an agent when the steps genuinely depend on what you discover partway through — "find why this test is flaky and fix it" — not when you already know the steps and just want it to feel advanced.

What agents are not

An agent isn't conscious, isn't pursuing goals of its own, and doesn't keep running when you're not looking. It has exactly the capabilities you handed it and nothing more — if you didn't give it a tool to send email, it cannot send email, no matter what it says.

It also isn't automatically better than a workflow. An agent pointed at a task with a clear fixed sequence will often do that task worse — more slowly, less predictably, and at higher cost — than five lines of ordinary code calling the model once per step.