TL;DR:
Websites are built for people who can see the whole page. Agents often see only part of it.
Common blockers: cookie banners, pop-ups, sign-in walls, and layouts that changed since last time.
A good agent retries a few times, then hands the task back and says what it saw.
You ask an AI agent to book a table or buy a ticket. It opens the site, finds the button, clicks it, and nothing happens. A cookie banner is sitting on top of the button, and the click landed on the banner. To you the problem is obvious in a glance. To the agent it is a puzzle.
This article explains why websites trip up agents, walks through the cookie banner case, and shows what a well behaved agent does when it is stuck.
The short answer: sites are made for eyes and hands
A web page is designed for a person who sees all of it at once, understands what is in front, and closes what is in the way. An agent works through a narrower window. It reads the page's code, or looks at a picture of the screen, or both. Either can be fooled by something that sits on top of what it wants, or that moved since the last visit.
What happens step by step when an agent uses a page
It reads the task: what to do and where.
It looks at the page and decides which element to use, such as a button labeled Book.
It acts: click, type, scroll.
It checks what changed. Did the page move on, show a message, or stay the same?
If the result is not what it expected, it tries again or tries another way.
Almost every failure is in step two or step four. Either it picked the wrong thing, or it did not notice that nothing had changed.
Walk-through: a cookie banner covering a button
Here is an example with a made-up site. The task is to click Reserve on a restaurant page.
The agent loads the page and finds a button called Reserve near the bottom.
A banner across the bottom says the site uses cookies, with two buttons: Accept and Manage.
The agent clicks where Reserve is. The click lands on the banner, because the banner is in front. Nothing happens.
It checks the page and sees no change. That is the clue that something is wrong.
On a retry it looks again, notices the banner, and decides to deal with it first.
Here the choice matters. Accept and Manage change your privacy settings, so a careful agent does not decide that for you. It can ask you which you want, or choose the least intrusive option you have told it to prefer.
With the banner gone, Reserve is clickable and the task goes on.
The agent was not stupid. It lacked a habit that people have: when a click does nothing, look for what is in the way.
The common blockers
Cookie and consent banners. They sit on top of content and need a decision.
Pop-ups and overlays. Newsletter requests, chat bubbles and offers that cover the page.
Sign-in walls. The page you want needs an account.
Human checks. Puzzles meant to confirm you are a person. An honest agent does not try to beat them.
Changing layouts. A site moves a button or renames it, and what worked last week no longer fits.
Slow or partial loading. The button exists in the code but is not ready yet.
Dates, dropdowns and custom widgets. These are built in unusual ways that are hard to operate.
Why retries help, and when they stop helping
A retry is a fresh look. After one failed click the agent looks again at the page, which may now show a message or a changed layout. Two or three retries with a new approach each time are reasonable. The same click ten times is not, because nothing about the situation has changed.
Good agents change something each time: wait for loading, scroll the element into view, close what is in front, or look for another route to the same result. If none of that works, the right move is to stop.
Handing back instead of guessing
When an agent is stuck, the worst thing it can do is guess, because a guess on a real website can mean a purchase or a message you did not want. The better move is to stop, say what it saw, and let you finish or choose. A useful hand-back says three things: what it was trying to do, what is in the way, and what you could do.
Note: Agent features that work inside your apps and browser are new. DuckyHelper's are in early access. It's built to ask first before it sends a message or pays for anything, and not to type your passwords or card numbers.
What it can't do, and what stays with you
It cannot be sure a page looks the same tomorrow.
It should not decide your privacy choices without telling you.
It should not get past a check made to stop automated programs.
It cannot know what you meant if the request was vague. Be specific about the date, the place and the number of people.
Privacy and control
Before you let an agent act on a site, think about what the page can see. If you are signed in, the agent acts as you. That is powerful and it is why you should watch the first few runs. Prefer tools that show you each action, let you stop at any time, and ask before anything irreversible.
Try it yourself
Pick a low-stakes task on a site you know, such as finding opening hours.
Watch the agent. Note the first time it gets stuck and what was in the way.
Check that when it stops, it tells you why in plain words.
If it guesses instead of asking, do not give it a task with consequences yet.
Frequently asked questions
Why do agents struggle more than people with banners?
People see the whole screen and instantly spot what is in the way. An agent may work from the page's code or a screenshot and needs to notice that its click did nothing.
Is it safe to let an agent accept cookies for me?
Only if you have said that is what you want. Consent choices are yours, so a careful agent asks or follows a setting you chose.
Get started
