Two years ago, the line was simple: an assistant answers, an agent acts. That line no longer holds. ChatGPT's agent mode opens a browser, fills out forms and hands you back a finished spreadsheet. Tools like ChatGPT Work or Claude's Cowork app go further: you describe an outcome, they work on your files and apps on their own, then return a finished document. Your assistant already knows how to click, browse and run code. If "acting" is no longer enough to define an agent, then what is?
It is a real question, and not only for beginners. The right answer is not about what the tool can do, but about who holds the wheel. An assistant acts inside a conversation that you steer. An AI agent, by contrast, takes the wheel: you set an objective, and it runs the task from start to finish. The rest of this article unpacks that idea level by level, then shows you how to begin safely.
The line has moved: from action to autonomy
Forget the idea that an agent is "the AI that does things". Today, almost everything does things. The real question is: who decides the next step?
With an assistant, even a well equipped one, you stay in the loop. It acts, comes back to you, you follow up, you correct course. You approve every turn. That is how the chatbots and assistants you already use work.
With an agent, you step out of the loop. You give a goal, not a procedure. The agent breaks the goal into steps, acts, checks the result, corrects, and starts again until the mission is done. It decides the next move itself.
Autonomy, then, is not a switch but a dial. An assistant sits at the low end, a deployed agent at the top, and there is a whole territory in between. That is where the real story plays out.
The building blocks of an agent
To place a tool on that dial, it helps to know the blocks that make up an agent. An assistant already has some of them. An agent pushes them all, and the last one above all.
A goal. You give an outcome, not a list of instructions. "Find three time slots I share with this client next week" rather than "open the calendar, click here, compare there".
A plan. The agent breaks that goal into steps. This is the reasoning, handled by a language model that thinks through the path to follow.
Tools and actions. Browsing, reading an email, filling a spreadsheet, querying a database, triggering an app. This is what your assistant can already do, so it is no longer the thing that sets an agent apart.
A loop. The agent tries, observes, corrects, starts again, across several steps, without you taking back control between each one.
A memory. It keeps track of what it has already done and learned, so it can carry a long task.
A degree of autonomy. This is the decisive block. How far does the agent go on its own before handing control back to you? That setting, more than anything else, is what separates an assistant from an agent, and one agent from another.
Three levels of autonomy, from copilot to deployed agent
The simplest approach is to think in tiers. The same artificial intelligence can be wired as an assistant or as an agent: what changes is the length of the leash.
| Level | Who drives the loop | Examples | When to use it |
|---|---|---|---|
| 1. The equipped assistant | You, turn by turn | ChatGPT, Claude in conversation | Exploring, producing one-offs |
| 2. Agent mode | You set the objective, it runs a supervised session | ChatGPT Work, Claude's Cowork | Delegating a one-off multi-step task |
| 3. The deployed agent | The agent, continuously, triggered by an event | Lindy, Sierra, Devin | Automating a recurring, connected process |
At level 1, you already know the drill. The assistant acts, but inside a conversation you lead turn by turn. You approve each step.
At level 2, you hand over an objective and the tool opens a workspace, chains the steps on its own for a few minutes to half an hour, then returns the result. It is already an agent, but kept on a short leash: it stays supervised, asks for your go-ahead before sensitive actions, and works by session. The task starts, runs, ends. Nothing watches your inbox while you are away.
At level 3, the agent runs continuously and triggers itself on an event: an incoming email, a customer ticket, a deadline. It is connected to your systems and moves with little supervision. You no longer open it to use it: it runs in the background and lets you know when it needs you.
What it is actually good for
The clearest examples are the everyday ones. An agent can watch your inbox and draft replies while you do something else. It can answer your customers around the clock, on chat as well as on WhatsApp. It can pick up the same report every week from your data, without you having to prompt it. And, for technical teams, it can take a coding task and carry it through to testing.
You do not need to memorize the names of these tools. What matters is the click: unlike an assistant, an agent stays plugged into a process and holds the loop without you. When one of these uses matches a real need, the place to look is the AI agents category.
How to get started, one level at a time
Good news: you do not have to jump straight to level 3. The best way to learn is to climb the tiers.
First, stay at level 1, but do it better. Most people underuse their assistant. Give it context, let it browse and run things, iterate instead of settling for the first answer. You will already see much of the agentic logic at work.
Move up to level 2 on a real task. Turn on your assistant's agent mode (ChatGPT Work, Claude's Cowork app) and hand it a concrete multi-step mission: compare offers and fill a table, reorganize a folder of files, prepare a report. Watch how it chains the steps, and above all where it still gets things wrong. This is the most instructive stage.
Approach level 3 with care. When a process comes back every week and is connected to your tools, try a dedicated agent. Start with a narrow, reversible scope, with mandatory approval before any irreversible action.
The precautions: the higher you climb, the more the leash matters
Every tier gained in autonomy is paid for in vigilance. A few reflexes will spare you unpleasant surprises.
An error compounds. Because the agent works in several steps, an early mistake spreads and skews the rest. The longer and more autonomous the task, the more the checkpoints matter.
Mind the permissions. An agent with access to your mailbox, your files or a payment method can do a lot, including what you did not intend. Delegate on a leash, not blindfolded: read-only or mandatory approval until trust is earned.
Costs can climb. Each step consumes resources, and a session that runs longer than expected costs more than a question to an assistant. Set a boundary.
Check the result. An agent sometimes gets things wrong with confidence. A well presented deliverable is not necessarily a correct one. Keep the final say on what matters.
Protect your data. Before plugging an agent into sensitive or confidential information, check what the tool does with it and where it travels.
An agent is not an AI that acts, it is an AI you hand the loop to
Here is what to keep. Since every tool now acts, it is no longer action that makes the agent, but autonomy: who gets to decide the next step. An assistant acts under your direction, an agent directs itself toward the objective you have set.
The right way to begin is not to aim for the most autonomous agent right away, but to climb the levels: get the most out of the assistant you already know, test an agent mode on a real task, then deploy a dedicated agent when a process is worth it.
