Guardrails
Where a run stops on its own, before it could spend money, destroy something or get past a security check, and why.
The agent acts on your live site, so some things it must never do: pay for anything, delete anything, close an account, or type a two-factor code. It's told not to, but a model can be talked into things by the page it's reading. So the worker checks every page and every action itself, and ends the run when one crosses the line.
A run ended this way has the outcome Stopped for safety. Its report says "A safety rule ended
this run, so it has no score," followed by the reason, such as Clicking "Delete project" was stopped because it could destroy something.
Pages that end a run
Every time the page changes, before the agent sees it, the run stops if the page:
- has a payment form from a payment provider such as Stripe, PayPal, Braintree, Adyen, Square, Checkout.com, Paddle, Lemon Squeezy, Klarna or Shopify, even inside a frame
- asks for card details
- asks for a two-factor or one-time code
Actions that end a run
Before every action, the run stops if it's about to:
- go to an address whose path includes checkout, purchase, payment, pay, delete, delete-account, close-account or cancel-subscription
- click something that would delete an account, spend money or destroy something
- choose an option that would do any of those, or that picks a way to pay, like a credit card or PayPal
- press Enter or the up or down arrow on something that would do any of those
"Spend money" covers words like pay, buy, purchase, place order, checkout, subscribe and upgrade. "Destroy something" covers words like delete, remove, erase, revoke, deactivate, unsubscribe, cancel a subscription or order, reset all data, and transfer ownership.
How an action is judged
A click is judged by the real element on the page, never by how the agent describes it. The agent could call a button "Next" when it says "Delete account", and a page can try to steer what the agent says. So the worker reads the element's own text, and the text of any form or dialog around it. That's how a harmless-looking "Yes" in a dialog asking "Delete this workspace?" still stops the run.
A choice in a dropdown is judged the same way, plus the option itself. A dropdown that offers a way to pay stops the run whichever option is chosen.
A key press acts on whatever has focus, so Enter or an arrow is judged as a click on that element: Enter in a field submits its form, the same as clicking the form's button. Escape and Tab only close things and move focus, so they always go ahead. With nothing focused, a key goes ahead. With focus inside a frame, which the worker can't read, the key isn't pressed and the agent is told so, but the run carries on.
It errs towards stopping
The patterns match English words only, and they're deliberately broad. Expect some stops that look unnecessary: a "Remove filter" button, an "Upgrade" link in a menu, or any button inside a form that mentions deleting something. Stopping a run that would have been fine costs you one run. Not stopping one that wasn't fine could cost you a real order or real data.
On a site in another language, the patterns won't recognise the words, so the page and address checks do more of the work. That's another reason to use a test account with nothing at stake.
Why a stopped run has no score
A guardrail ended the run on purpose, so its number would say nothing about your site. The report still shows what happened before it stopped: the steps, the friction points and the recommended changes. The scorer is told the stop wasn't the site's fault or the visitor giving up.
The cost cap also ends a run with Stopped for safety. See how a run ends.
Writing goals around them
If a goal keeps stopping at the same place, that place is probably payment or deletion. Write the goal to end just before it, such as "Get to the point of paying for the Pro plan," and use success criteria the agent can see before that step. If it still stops there, the steps in its report show how the visitor got that far.