What should an AI agent show you before it acts?

Key points

• An approval only protects you if you can check it, and you can't check a sentence the AI wrote about itself.
• A good approval shows four kinds of real thing: a picture, one number, a strip drawn to scale, or the document on paper.
• On nox.cool's cards, the title, price, picture and button come from the checkout, the payment or the page, never from the AI's words.
• If the server can't stand behind a number, the number doesn't go on the card.
• Spending money takes a press held for most of a second. Everything else takes one press.
• Sign-in codes and hidden text are dealt with before they reach the AI or the card.

Before an AI agent books, pays or sends, it should show you the thing itself: the real page, the actual price, the email as it will arrive. A sentence describing the action isn't a check. It's the AI's own account of what it's about to do.

We rebuilt how Nox, the agent on nox.cool, asks for your OK. The old cards looked like any chatbot's reply, and the problem was bigger than looks. A chatbot describes things. An agent is about to do them.

#Why is a yes or no from an AI agent so easy to get wrong?

Because you're asked often, you're busy, and the card looks the same every time. By the tenth "Nox wants to click Continue on example.com", you've stopped reading. The approval is still there, but it has stopped protecting you.

The usual fix is to ask less. That helps, and nox.cool lets you allow things you're happy with, like emails to one person. But it leaves the harder question: when the agent does stop and ask, what should you be looking at?

#What's wrong with a sentence that describes the action?

The sentence comes from the thing you're checking.

Say an agent has misread a hotel page and is about to press a button for a different room, or for a paid upgrade it took for the free one. If you ask it to describe the action, it will describe what it believes it's doing. The summary carries the same mistake, written confidently, and you approve the mistake.

So we set one rule for these cards:

The card exists to check the AI, so the AI doesn't get to write it.

On nox.cool's OK cards, the title, the price, the picture and the button come from the checkout, the payment or the page. If the server can't stand behind a number, the number doesn't go on the card.

#What should an approval show instead?

We ended up with four kinds of thing, and nothing else. If something isn't one of these, it doesn't go on the card.

A real picture. When Nox is about to press a button on a website, you get a picture of that page with the button ringed. Pictures stay black and white until you reach for them, so a card doesn't shout, and passwords and card numbers are masked out of the picture.

One number. The amount, large. On a website, the figure has to be on the page itself. We'd rather show no number than one we can't back up.

A strip drawn to scale. Dates are easier to see than to read. A hotel stay is a row of days with your nights in black, and a mark on the last day you can cancel, when the page gives one.

The document on paper. An email appears on a white sheet, even in dark mode, because it's leaving your app and going to someone else. You see the recipient, the subject, the body and your signature, as they'll be read.

What we left out matters as much. We tried chips, icon tiles, arrows and stamps a few weeks ago, on a different card, and it read as busy and amateur. Decoration makes a card look designed. It doesn't help you decide.

#How does it work when Nox books or pays?

Here's what happens when you ask Nox to book a hotel:

1. Nox searches, compares and gets to the last step, on its own computer in the cloud.
2. Before the final press, it stops. The server takes a picture of the page around that button and reads the total from the page.
3. A card appears in your conversation, with the page, the price, the nights and the payment card that would be used.
4. Nox waits. It does nothing else on that task until you decide.
5. You hold the button for most of a second, and the booking goes through. Or you press Not now, and it asks what to change.

If you've turned on Touch ID or Face ID for money, it asks for that too.

#How does it work for an email?

The words of an email are the thing you're checking, so the card shows them whole, as they'll arrive. The button names the person: Send to Anna, not "Approve". If a word is wrong, press Edit and fix it there.

#Why does shopping need its own shelf?

Choosing is a different job from approving. When you're choosing, you want to compare, so a written list of products makes you do the hard part in your head.

On nox.cool, products sit side by side with their photos, prices, ratings and whether they're in stock. The one Nox would pick is marked, and it says why in one line, like "The cheapest that's fully waterproof". That line is the AI's opinion, which is fine: it's advice on a shelf, not a check on the AI. The check comes later, when you OK the total.

#What does this mean beyond approvals?

The same thinking applies anywhere you need to see where you stand. The Wallet tab beside a task used to be a long list of amounts. Now one strip shows where your money is, drawn to true proportions, by wallet or by coin. Open trades sit together, with the one closest to liquidation first.

Projects work the same way. Click one in the sidebar, and the agents on it appear underneath, with the lead marked.

Plain text is still right for plenty of replies. If Nox is answering a question, a paragraph is the best shape for it. We only change the shape when you have to look at something or decide something.

#What about pages that try to trick the agent?

A picture helps here too. Web pages can carry instructions aimed at AI agents, written in text that people can't see. Nox treats everything it reads as information and never as orders, and it still pays to have more than one line of defence:

• You see the real button, on the real page, before it's pressed.
• Hidden text is shown up. Invisible characters in anything you're asked to OK appear as a visible mark, so what you approve is what's there.
• Sign-in codes stay out of the AI's sight. When Nox reads your Gmail or Outlook, sign-in codes and one-time sign-in links are taken out first. If it needs one to finish a job, it asks you, and uses it without ever reading it.
• Old sign-ins end by themselves. A browser you've stopped using is signed out after 14 days.

#How do I try it?

1. Sign in to nox.cool. It's in early access, by invite for now.
2. Ask for something with a final step: "Find me a hotel in Lisbon for next weekend, and ask me before you book."
3. When the card appears, look at the page, the price and the dates. Then decide.

For the full list of what's new, see the announcement.

#Questions and answers

What is an approval in an AI agent?

It's the point where an AI agent stops and waits for a person before doing something that matters, such as paying, booking or sending a message. On nox.cool it's a card in the conversation, and the agent is paused until you decide.

Why isn't a text description enough?

Because the description comes from the AI being checked. If the AI has misread a page, its summary carries the same mistake, and you approve the mistake. Showing the page itself lets you see what the AI may have missed.

Who writes what's on a nox.cool OK card?

The title, price, picture and button come from the checkout, the payment or the page. An email shows the words that will be sent, because those are the thing you're checking. Actions with no picture or price yet keep a plain card.

What if a web page tries to trick the agent?

The agent treats what it reads as information, never instructions. Invisible characters in anything you're asked to OK appear as a visible mark, and a picture of the real page lets you see the button that will be pressed.

Can I stop being asked?

Yes, for some things. You can always allow a specific kind of action, like emails to one person, right from the card.

Is this available now?

Yes, in nox.cool's early access, which is by invite for now. The web app and the Mac app both show the new cards.