Welcome! Let's untangle a buzzword
You've probably heard the word "agent" a lot lately. It's in product launches and in news headlines, and sometimes it pops up in your own apps. So what does it actually mean? And how is an agent different from the chatbot or assistant you already use?
That's our job today. By the end, you'll have a simple mental map, and you'll be able to put tools you know onto it. You'll also learn to raise an eyebrow when someone uses the word too loosely.
Start with a simple question: who does the work?
Here's the easiest way in. When you use an AI tool, ask yourself one thing: who is doing the steps, you or the AI?
Picture three helpers in a kitchen.
The first helper is a cookbook that talks back. You ask, "How do I make a risotto?" and it tells you. That's useful, but you do all the cooking.
The second is a sous-chef standing next to you. You say, "Chop these onions," and it chops them. Then you say, "Now stir the rice," and it stirs. It helps with real tasks, but you direct every step.
The third is a chef you hand a goal to. You say, "Make dinner for four, nothing with nuts, ready by seven." The chef plans the menu, checks the cupboards, cooks, and tells you if something's missing.
That's the whole idea. Chatbot, assistant, agent.
The chatbot: it answers
A chatbot is built for conversation. You type a message and it replies. It's great for explaining things, drafting text or brainstorming ideas.
But notice what it doesn't do. It doesn't go off and do anything in the world. Everything happens inside the chat window, and you take the answer and act on it yourself.
The assistant: it helps you do things, step by step
An assistant goes a bit further. It can often use a few tools: searching the web, reading a file you upload, summarising your inbox, drafting a reply in your email app.
Here's the key thing, though: you're still in the driver's seat for each step. You ask, it helps, then you ask again. Most of the AI tools you use day to day spend most of their time here.
The agent: it works towards a goal
An agent is what you get when you hand over a goal rather than a single request. It works out the steps, uses tools to carry them out, and keeps going with some independence until it's done or needs you.
Here's a real example. OpenAI's help pages describe ChatGPT's agent mode as able to navigate websites, fill out forms and edit spreadsheets on your behalf, and they say tasks usually take 5 to 30 minutes. That's very different from a two-second chat reply! Anthropic's Claude in Chrome extension can read, click and navigate websites. In 2025, Microsoft announced an "Agent Mode" in Copilot for Word and Excel that plans and carries out multi-step edits.
So what do these have in common? You describe the outcome, and the AI does a chain of actions to get there.
Why it's a spectrum, not a box
Now here's the twist. These three aren't separate boxes with hard walls. They're points along a spectrum, a sliding scale running from "just answers" to "acts on its own towards a goal."
How do we know? Because one product can sit at different points depending on how you use it. ChatGPT, Gemini, Claude and Copilot can all behave like a plain chatbot one minute and switch into an agent-style mode the next. So when you ask "Is this tool an agent?", the honest answer is usually: "It depends on which feature you're using, and how much it's allowed to do by itself."
Let's try it. Think of three AI tools you've used this month. For each one, ask:
- Does it only answer me?
- Does it help with tasks while I direct each step?
- Or can I give it a goal and let it carry out several steps on its own?
A spell-checker with AI suggestions? That's near the chatbot-and-assistant end. An email app that drafts a reply when you click a button? That's an assistant. A tool you can tell "research these five suppliers and put the results in a spreadsheet" before you walk away? Now you're at the agent end.
There's no wrong answer here. The point is to get into the habit of asking.
Spotting "agent washing"
Because "agent" is a hot word, it gets stuck on things that don't really earn it.
The research firm Gartner has a name for this: "agent washing." In a June 2025 press release, Gartner described it as rebranding existing products, like AI assistants, chatbots and older automation tools, as agents without substantial agentic capabilities. It estimated that only about 130 of the thousands of vendors claiming to sell "agentic AI" were real. Keep in mind that's one firm's estimate from one point in time, not a precise count. Still, it's a useful warning.
So how do you spot it yourself? When you see "AI agent" in an ad, ask a few plain questions. What can it actually do, not just say? Can I give it a goal, or only one request at a time? Does it take any steps without me prompting each one?
If the honest answer is "it chats and suggests things," you're looking at a chatbot or assistant in a fancy coat. That's fine! It can still be very useful. It just isn't what the label promises.
Key takeaways
- A chatbot answers, an assistant helps with tasks while you direct each step, and an agent takes a goal and carries out several steps with some independence.
- The quickest test: who is doing the steps, you or the AI?
- It's a spectrum, not a set of boxes. The same product can act as a chatbot or as an agent depending on the feature and how much it's allowed to do.
- "Agent washing" (Gartner's term, June 2025) means calling a product an agent when it lacks real agentic abilities. Ask what it can actually do on its own.