TL;DR:
Some steps are meant for a human: checks that you are a person, passwords, and approvals for sending or paying.
A good agent stops at these steps, says why, and waits for you.
Handing back protects your accounts and your money.
An agent is partway through a task when a page asks it to pick all the pictures with traffic lights. It stops and says: I need you to do this part. That can feel like the agent giving up. It is the agent working correctly.
This article explains the three main moments when an agent should hand control back, walks through a CAPTCHA, and describes what you should expect to do.
The short answer: some steps belong to you
Three kinds of steps should not be done silently by software acting for you.
Human checks. Puzzles and tests built to confirm a person is present.
Sign-ins. Your passwords and card numbers. These prove who you are.
Approvals. Sending a message, paying, deleting, or anything you cannot take back.
In each case the whole point of the step is that you are the one doing it.
What happens step by step
The agent works through the task and reaches a step it should not do alone.
It stops and does not guess or try to get around the step.
It tells you what it was doing, which step it reached, and what you need to do.
You complete the step. For a CAPTCHA, you solve it. For a sign-in, you type your own password.
You tell the agent to continue, or it notices the page moved on, and the task resumes.
Walk-through: a captcha in the middle of a task
This is an example. You ask an agent to download your monthly statement from a utility's website.
It opens the site and signs in, using a session you already started. Nothing has been asked of you so far.
On the statements page it clicks Download. The site shows a puzzle: select all the squares containing a bicycle.
The agent stops. It says: The site wants to check that you are a person. I have not tried to solve it. Please complete the puzzle and tell me when you are done.
You solve the puzzle. The page shows the statement.
You say: Done. The agent saves the file, tells you where it is, and finishes.
The agent did nearly all of the work. It handed back one step that took you a few seconds. The alternative, an agent that tries to defeat such checks, would be breaking the site's rules and would put your account at risk.
Notice that the agent said what it had not done. It did not claim success and it did not try the puzzle once to see. Plain reporting of where the task stands is what lets you trust the steps it did finish.
Sign-ins and card numbers
A password and a card number are the keys to your accounts and your money. An agent that types them is an agent that has them. A safer design never types them for you. It stops at the sign-in page, you enter your details yourself, and the agent carries on from there.
Often you will not need to sign in at all. If you are already signed in in that app or browser, the agent can use the existing session. That is why you should be careful which accounts are open when you start an agent on a task.
Approvals: when sending and paying need your yes
Some actions cannot be undone. A sent email stays sent. A payment leaves your account. Deleting a file may be permanent. For these, the agent should show you what it is about to do, in full, and wait for a yes.
A good approval shows the details: for an email, the recipient and the text; for a purchase, the item, the price and the total. If you cannot see those, you cannot meaningfully approve.
Note: DuckyHelper's agent features 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
It cannot decide for you whether a payment is a good idea.
It cannot sign in for you without your involvement.
It cannot always tell when a page has a hidden check. It may fail first and then ask.
It cannot promise that every site will allow automation. Some sites say no in their terms.
Privacy and control
Control is the reason for all of this. Look for a tool that gives you a way to stop it at any time, shows you what it did, and keeps the risky steps in your hands. Read what the tool says about what it stores. If you hand over a task that involves other people's information, such as customers or classmates, think about whether you are allowed to.
Try it yourself
Give an agent a harmless task that needs a sign-in or a check, such as downloading a receipt.
Watch what it does at the sign-in. It should stop and ask.
After you complete the step, see if it continues without repeating work.
If it asks for your password in the chat, stop. Do not paste it.
Frequently asked questions
Does handing back mean the agent failed?
No. It means it recognised a step meant for you. Finishing the other steps and asking for one is better than guessing.
Can I tell it to solve CAPTCHAs for me?
A responsible agent should not. They exist to confirm a person is there, and getting around them may break a site's rules.
Get started
