What is an AI agent? (in plain English)

“Agent” is the word of the year, and like most words of the year, it’s used to mean ten different things. Let’s fix that. The concept underneath is simple, and once it clicks, a lot of what’s happening in AI right now suddenly makes sense.

Start with what an agent is not

You already know chatbots. You type a message, the AI types back, and then it waits. It’s a brilliant conversation partner, but it’s fundamentally reactive: one turn, then it stops. If the task needs five steps, you drive all five — you ask, read, decide what’s next, ask again.

An agent flips that. You give it a goal, and it drives the steps itself.

The loop that makes an agent an agent

Here’s the entire idea in three words: act, observe, decide.

An agent:

1. Acts — takes a concrete step toward the goal (reads a file, runs a command, searches the web, sends a request).
2. Observes — looks at what actually happened (the file’s contents, the command’s output, the error message, the search results).
3. Decides — based on what it observed, chooses the next step. Then it loops back to act again.

It keeps cycling through that loop — acting, observing, deciding — until the goal is met or it hits something only you can resolve. That loop is the whole difference between an agent and a chatbot. A chatbot answers; an agent gets something done.

A human analogy: a chatbot is like texting a knowledgeable friend one question at a time. An agent is like handing a capable assistant a task and saying “sort this out” — they go off, try things, notice what works, adjust, and come back when it’s handled.

Why the loop is such a big deal

Because most real work isn’t one step. It’s “find out why the report is wrong” — which means looking at the data, forming a hypothesis, checking it, being wrong, forming another, and finally finding the broken formula. A chatbot can advise you through that. An agent can do it, because it can act on what it learns at each step instead of waiting for you to relay it.

This is also why agents can surprise you in good ways. Because they observe real results rather than guessing, they can recover from a wrong first move — see the error, understand it, and try something else — the way a person would.

Claude Code is an agent you can watch

The clearest way to understand agents is to see one work. Claude Code is an agent that lives in your terminal:

• You give it a goal: “the login page is broken — find out why and fix it.”
• It acts — opens the relevant files, runs the app, reproduces the bug.
• It observes — reads the error, sees what’s actually happening.
• It decides — forms a theory, makes a change, runs it again to check.
• It loops until the page works, then shows you exactly what it changed.

You’re not relaying information between steps. You set the goal and review the result. That’s the agent shape.

What gives an agent its reach: tools

An agent is only as capable as the actions it can take. A model that can only produce text isn’t much of an agent — it has no hands. What turns it into one is tools: the ability to read and write files, run commands, search the web, query a database, click around a web page.

This is why standards like MCP matter so much right now. Every tool you connect is a new kind of action the agent can take — a new way to act and a new thing to observe. Connect a browser and it can check whether your site loads. Connect your data and it can investigate a number that looks wrong. More tools, more reach.

The thing everyone worries about: control

If an agent takes actions on its own, what stops it from doing something you didn’t want? Good agents are built so that autonomy is bounded, not unlimited:

• Permission prompts before consequential or irreversible actions — the agent proposes, you approve.
• Scoped access — it can only touch the files, tools, and data you’ve allowed.
• A human in the loop for high-stakes decisions, with the agent drafting and you deciding.

The goal isn’t an agent that does everything unsupervised. It’s one that handles the legwork autonomously while keeping you in control of the moments that matter. Done well, you get the speed of delegation with the safety of oversight.

Why this is the direction everything is heading

The reason “agent” is everywhere is that it’s the natural next step once models got good enough to be trusted with a loop. A model that can reason well is useful as a chatbot. A model that can reason well and act on the results of its own actions is useful as a teammate. As the models keep improving, the value shifts from “give me a smart answer” to “go handle this” — and “go handle this” is the agent’s whole job.

Understanding the act-observe-decide loop is the key that unlocks the rest. Skills, tools, MCP, subagents — they’re all ways of giving that loop more reach and more reliability. Once you see the loop, you see the pattern everywhere.