Why did the assistant stop and ask me to take over?

Sometimes the assistant gets part of the way through a job in the browser, stops, and asks you to do the next bit yourself. That is deliberate. There is a short list of things it will not do on your behalf, and when it reaches one of them it stops rather than guessing.

Two things are true every time this happens:

• The browser stays open on the page it reached. Nothing is thrown away, and you do not start the task again from the beginning.
• It is not a failure. The job is parked, waiting for you, and the assistant tells you what it stopped at.

A password, a passcode or a security code

The assistant does not type credentials. If a task reaches a password box, a PIN, or a one-time or two-step code, it stops there and asks you to type it yourself in the live view.

The same goes for payment card numbers. And if it cannot tell what a field is for, it will not type into it at all rather than guess.

A bot check

The assistant never attempts a CAPTCHA or a "confirm you're human" step. When a page puts one in the way, the run parks, the browser stays open on that page, and you are asked to clear the check yourself and say when to carry on.

A site that wants you to sign in

The assistant does not sign in as you. When a task needs an account, it proposes a sign-in and you tap Approve before any browser opens for it. It then opens the browser on that site's sign-in page and waits while you type your own details and any two-step code. It is not driving at all during that window. The panel says so above the box:

"Sign in below and the session is saved for future runs. This window isn’t recorded, and your password is never stored — only the signed-in session is kept."

Press I’m signed in when you are through, and later runs on that site start already signed in.

Saved sign-ins do go stale on their own — after a password change, a "log out of all devices", or simply an expired session. When a run finds itself logged out it says so and asks you to sign in once more from the panel, rather than retrying blindly.

Anything that would actually go through

Before an action that would commit you to something — a purchase, a payment, an order placed, a form submitted, a message sent, something published or deleted — the assistant stops and asks you to confirm, unless your own request asked for that specific thing. It tells you what it is about to do first.

That stop applies to each action on its own, and it covers the keyboard as well as the mouse. If pressing Enter would submit a form, the assistant fills the form in and stops short of submitting it.

What to do: if it is what you wanted, tell it to go ahead. If it is not, finish the step yourself in the panel, or tell it what to change first.

A site outside the ones you named

If you scoped a task to particular sites — "only look on these two" — that restriction is enforced, not merely noted. Going anywhere else is refused, including a page the site itself redirects to, a link that opens a new tab, or a switch to a tab that is already open. The assistant tells you which sites were allowed and which one it stopped at, and offers to add it if that is what you want.

It also sticks to ordinary web pages. Links that are not ordinary web addresses, and addresses pointing at a private or internal machine, are refused.

A page trying to give the assistant instructions

Some pages carry text written to be read by an AI assistant rather than by you. Anything on a page is treated as page content, never as a request from you. The assistant does not act on it, and it tells you it saw it.

The run reached its spending limit

A single browser run has a ceiling on what it can spend. A long job can reach it before it is finished, and the run stops there rather than carrying on. The browser is left open on the page it reached, and the assistant tells you what still needs doing — ask it to carry on and it picks up in the same browser.

When it hands over: what is on screen

You do not have to grab anything. When the assistant stops for one of these, the browser is already yours. The status chip in the panel header reads You're driving, and the header button reads Hand back rather than Take control.

1. Do the step yourself in the panel — type the code, clear the bot check, confirm the purchase.
2. Press Hand back. A short message goes into the chat on your behalf, so the assistant knows straight away, and you will see "Handed back — the assistant knows and is picking it up."

After a hand-back the assistant reads the page again before it does anything, so it carries on from where you left it rather than starting the task over.

Taking the controls when it has not stopped

You can also jump in part-way through. Press Take control in the panel header, or click anywhere on the live browser view — hovering over it shows "Click anywhere to take control".

The assistant stops immediately, you will see "You're driving. The assistant will wait.", and while you have the wheel it will not click, type or navigate. Ask it to do something in the browser meanwhile and it will say it will pick that up as soon as you hand back. The same button, now reading Hand back, returns the wheel. If the change does not go through you will see "Couldn't change who's driving" — try once more.

Rephrasing will not get round it

These stops sit underneath the assistant rather than being a matter of its judgement. Asking again in different words, or asking it to find another route to the same thing, will not work — it will stop in the same place. The quickest way through is always to do that one step yourself and tell it to carry on.

If a run ends after one of these stops, the assistant tells you what stopped it rather than reporting a half-finished job as done.

Tip: You do not have to take over the whole task. Clear the single step that stopped it, hand back, and the assistant picks up from what is on screen.