What Is an AI Agent, Really?

"Is ChatGPT a robot?" That is usually the first question people ask when the topic of AI agents comes up. The short answer is no. The slightly longer answer is that ChatGPT is a very clever text predictor, while an AI agent is something else entirely: a program that pursues a goal by taking actions in the world around it.

The confusion is understandable. Both are powered by large language models. Both can hold a conversation. But the difference between answering a question and acting on it is the difference between reading a map and driving the car. Once you see that distinction, the entire field of "AI agents" snaps into focus.

From Answering to Acting

A chatbot, at its core, is a conversion machine. You give it words, it gives words back. The interaction begins and ends in the chat window. An agent, by contrast, is built around a goal. You hand it an objective, and it must decide what steps to take, which resources to consult, and when to report back.

Consider a trivial example. Ask a pure chatbot to "send an email to my team about the meeting tomorrow." The chatbot will happily draft a perfectly formatted email. It might even paste it into the reply box. But it will not press send, because it has no concept of pressing send. Ask an agent to do the same thing, and it will draft the email, find the recipients in your contacts, open your mail client, and submit it.

The Mental Model

Here is the mental model I want you to keep. An agent is a loop with four moving parts:

A goal that defines success.
Perception — reading the current situation, from messages, files, or sensors.
Reasoning — deciding the next step using an LLM as the "brain."
Action — executing tools that change the world, then looping back to observe the result.

Notice what is missing: a human in the middle. The whole point of an agent is that it closes the loop itself. It looks at what it did, checks whether that moved it closer to the goal, and adjusts. That autonomy is what turns a passive tool into an agent.

"A chatbot tells you what could be done. An agent tries to do it, and tells you what happened."

Why the Distinction Matters

This is not academic nitpicking. The distinction changes what you can build and what you should expect. When you deploy a chatbot, you are buying convenience. When you deploy an agent, you are buying delegation. That difference carries real operational weight.

Think about what changes when software is given the authority to act:

It can run in the background while you sleep, not just when you are at the keyboard.
It can touch external systems — databases, APIs, calendars, payment rails — rather than staying trapped in a text box.
It accumulates a history of actions, giving you an audit trail of what it did and why.
It introduces new failure modes: it can act on bad information or take a wrong step before you notice.

That last point is the hidden cost of agency. The more power you hand over, the more careful you must be about guardrails, permissions, and oversight.

Not All Agents Look Alike

Agents come in many shapes. Some are single-step: they read a request, run one tool, and stop. Others are deeply autonomous, chaining dozens of calls over minutes or hours. Some live inside a chat interface and feel like particularly confident assistants. Others have no visible interface at all — they quietly monitor a queue, process tickets, and post results to a dashboard.

Here are a few concrete forms you are likely to meet:

Task agents that handle a specific job, like summarizing a codebase or triaging support tickets.
Workflow agents that orchestrate several steps in a fixed process, following a recipe rather than improvising.
Autonomous agents that decompose a high-level goal into sub-goals, choose tools, and self-correct over multiple iterations.
Multi-agent systems where several specialized agents collaborate, passing work to one another like a team.

They all share the same spine, though: goal in, action out, feedback loop in between. If you understand the loop, you understand every agent on this list.

A Working Definition

Let me give you a definition you can carry around. An AI agent is a system that uses a language model to decide among possible actions, executes an action through a tool, observes the result, and repeats this cycle until it reaches a goal or hits a stopping condition.

Each clause earns its place:

"Uses a language model to decide" — the brain is a generative model, not a lookup table.
"Decides among possible actions" — the model picks from a real choice set, not a single hardcoded path.
"Executes through a tool" — the output must leave the model and touch the world.
"Observes the result" — feedback feeds back in, closing the loop.
"Repeats until goal or stopping condition" — termination is part of the design, not an accident.

What an Agent Is Not

It helps to define the boundaries by negation. An agent is not just an LLM, because an LLM alone cannot act. It is not a rule-based script, because those do not reason about novel situations. And it is not a magic system that understands goals the way a human does — it is still a statistical machine, just one wired into an action loop.

Keep some healthy skepticism. The word "agent" is now a marketing favorite. A product that sends you a canned notification is not necessarily an agent. The test is simple: does it perceive, decide, act, and observe in a loop, without a human turning the crank each time?

Why Now

Agents are not a new idea; symbolic AI researchers dreamed of autonomous reasoning systems decades ago. What is new is the raw capability of modern language models. They make the "reasoning" step strong enough that the loop actually works in messy, real-world settings. Cheap inference, long context windows, and a rich ecosystem of APIs have turned a philosophical idea into an engineering practice.

Stepping Stones

If this is your first encounter, you now have the mental model. The next articles in this series will flesh out the pieces: how agents differ from plain chatbots, how the perceive-decide-act loop works in practice, what tools do for an agent, and why memory matters. Each builds on the loop drawn here.

For now, remember the one sentence that separates a tool from an agent: a chatbot hands you a suggestion, and an agent hands you a result. Everything else is detail.

The agent loop: goal, perceive, decide, act, observe.

Figure 1. The core loop shared by every agent. Each revolution moves the system closer to its goal.

Before you go measuring agents at work, pause and ask whether the thing in front of you actually closes the loop. If someone still has to press the button, it is a very good chatbot, and nothing more. If the system decides, acts, and checks its own work — that is where the real magic begins.

Start With a Goal, Not a Tool

When teams first try agents, they usually reverse the order. They pick a fancy tool first and hunt for a use case. Start instead with a job that involves real decisions and real consequences, where "typing an answer" is not enough. The goal will dictate the tools, the loop, and the guardrails. Get that part right, and the rest of the series will slot into place.