Why should an AI agent think twice about the same job?

Key points

• An AI agent normally works out a repeated job from scratch each run, and that's most of what the run costs.
• Nox can repeat a routine's browser steps without the main AI. In our test, about 11¢ a run became 0.03¢, and about 20 seconds became 5.
• It finds what to click by its name and role on the page, never by a position on the screen.
• If a step has gone, the AI takes over from that point with the whole record in front of it.
• Repeated steps go through the same checks as the AI's own, so anything that needs your OK still asks.
• It's off until you turn it on, because a repeat clicks what was clicked last time and doesn't choose.

nox.cool can now repeat a routine's browser steps without the main AI. In our test, a run that cost about 11¢ cost 0.03¢ and took 5 seconds in place of about 20. When a website changes, the AI takes over from that step, and anything that needs your OK still asks. It's off until you turn it on.

#Why does an AI agent's routine cost the same on day fifty?

Say you've asked an agent to check you in somewhere every Thursday: open the page, click the link, press the button, tell you it's done. The first run is impressive. It reads the page, works out the button, and gets it done.

The second run does all of that again, and so does the fiftieth. Each run starts fresh. The agent reads its instructions and everything it could do, looks at the page, decides, clicks, looks again and decides again. Most of what you pay for is that thinking, and on a job that hasn't changed, it's thinking about something it already knows.

We'd already made the clicking cheaper by handing it to a smaller, faster model. That helped a lot. The main AI still had to start up, read everything and run the job, though, so the bill stayed well above what a three-click job deserves.

#What does repeating without the AI do?

While a routine runs, Nox keeps a record of what it does in the browser. Then, once you've said so, the next run uses that record:

1. Nox opens the same address.
2. It looks at the page for what the last run clicked, and clicks it.
3. It carries on to the end of the steps.
4. A small AI model reads the last page, and writes you the result in a line.

The main AI never starts.

The saving comes from not asking an AI what to do when the answer hasn't changed.

Here's what we measured, on a test routine that opens a page, clicks its link and reports the heading it finds:

It's a small routine, picked because it's easy to check, and your own will vary. A routine with more pages costs more with the AI, and adds very little repeated.

#What happens when the website changes?

This is what usually goes wrong with recorded scripts. A button moves, and the script clicks the wrong thing or stops dead.

So Nox doesn't keep positions on the screen. It keeps what each thing is: a link called "Learn more", a button called "Book". On each run it looks for that on the page as it is now. A redesign that moves the button changes nothing.

If the button is gone, or two things fit, Nox doesn't guess. The repeat stops and the real AI takes over at that point, with the whole record in front of it and a note saying why it stopped. We tested this by renaming a step to something that wasn't on the page. The AI found the right link itself and finished the job, for 10.7¢, about what an ordinary run costs. The steps it took become the ones repeated next time.

So the worst case is an ordinary AI run. That's why it can be safe to leave running.

#Can a repeat spend money without asking?

No, and this is the part we cared about most.

A repeated step isn't a shortcut around the agent. It goes down the same path as a step the AI asks for. Your rules apply, and sites you've blocked stay blocked. A button that spends money or speaks for you stops for your OK.

We tested this with a recorded click on a button called "Submit order". The repeat stopped and showed the card, with the main AI nowhere in sight. Saying no ends that step, and the AI hears why.

Secrets stay out of it too. A run that typed into a password box, or used a saved login, is never kept as steps. Nor is one that typed something like a card number.

#When shouldn't you use it?

When the job involves choosing.

A repeat clicks what was clicked last time. If your routine is "find the cheapest flight to Lisbon", a repeat would click yesterday's cheapest flight, and that's a wrong answer delivered confidently. The AI should think those through each run.

That's why it's off until you say so. Nox offers it only once the AI has done a job the same way twice, and the offer says plainly: "It won't choose anything new." Runs that used anything beyond the browser, like email, connected apps or files, aren't offered it at all.

To turn it on, open Routines and click Repeat them without the AI, or ask Nox. To go back, click Think it through each run instead. Changing what the routine is meant to do clears its steps, so it starts thinking again.

#What other work shouldn't need an AI to wake up?

Once you start looking, there's more of this. Take a page showing your revenue. Waking an agent for every payment would cost real money to do something a counter could do.

So you can now ask: "Make me a page that shows my Stripe revenue as it comes in." Nox builds the page and puts it up once you've approved it. Then Stripe sends each payment straight to that page, with no AI involved. A refund takes off only what was refunded. The same payment sent twice counts once. A call that isn't really from Stripe, checked against a signing secret only you handle, is turned away.

Real tests earned their keep here. When we asked a real agent to build this page, it sensibly kept the figures private, and that showed us a page couldn't read private figures even for its owner: you'd have opened your own revenue page and seen nothing. So the Sites page now has Open it with your private data, which opens the page as you and only reads. The first version of the page also added refunds up instead of taking them off, because we hadn't told agents clearly enough which number to total. We fixed the instructions, asked again, and it came back right.

#Why point at things instead of describing them?

Telling an agent what's wrong in a document usually goes like this: "On page three, in the second paragraph, where it says thirty days…". It works, and it's slow, and it goes wrong the moment that phrase appears twice.

Now you can open the document and point. A PDF shows as its real pages. Select the words, click Comment, and say what should change. Or click Mark a part and draw a box round a table.

The agent gets exactly what you marked, with a little either side so it can find it again. In our test on a Word file, it came back with a new copy: "thirty" struck through, "sixty" tracked, nothing else changed, your original untouched.

Word files and decks show as their words for now, so you can select words in them but not draw on a slide. A note on a picture works too, including a screenshot: "this button, here". The agent sees the picture with your part boxed.

#What about the small things you just want on your screen?

Sometimes you don't want a website, only a timer. Ask for one and it stays beside the chat on every page until you close it. Drag it out of the way, or make it small and it keeps counting.

Each one is sealed in a frame on an address of its own, so it can't reach your accounts or anything else of yours. That's also why it needs no OK: only you see it.

#How do decisions stay decided?

If you work with several agents, you may have had this one: you settle something on Monday, and on Wednesday an agent suggests the opposite because it read something new.

Now you can say "we've decided on Stripe, not Paddle, that's final". You OK it on a card, and every agent reads it first on every task. We tried pushing against it with a real model: asked to set up Paddle, Nox pointed out what had been decided, said it wouldn't reopen that unless told to, and left the decision as it was. Taking a decision back needs your OK on a card too. No agent, routine or page can settle or unsettle anything without you.

#What does all this add up to?

Each of these comes from the same question: what actually needs an AI? Deciding does. Choosing does. Understanding a page you've never seen does. Clicking the same three things every Thursday doesn't, and neither does adding a payment to a total.

Nox still thinks whenever thinking is what the job needs. It just no longer has to think about the things it already knows.

If you'd like to try it, you can ask for early access at nox.cool. The short version of what's new is in the announcement.

#Questions and answers

Why does a daily AI agent routine cost so much?

Each run starts fresh. The agent reads its instructions and every tool it could use, looks at each page, and decides what to do next, step after step, even when the job hasn't changed since yesterday.

How much does repeating a routine save on nox.cool?

In our test, a routine that opens a page, clicks a link and reports a heading cost 13.2¢ and 10.7¢ with the AI, and 0.03¢ repeated. It took 26 and 18 seconds with the AI, and about 5 seconds repeated. Your own routines will vary with what they do.

What happens when a website changes?

Nox looks for each thing by its role and name. If it isn't there, or more than one thing fits, the repeat stops guessing and the AI takes over from that step. It then finishes the job, and the steps it took become the ones repeated next time.

Can a repeated routine spend money without asking?

No. A repeated step goes through the same checks as a step the AI takes. In our test, a repeated click on a button called Submit order stopped and showed an OK card without the main AI starting.

Is this the same as a browser macro or a recorded script?

It's similar in that it replays steps. It differs in what happens around them: your rules and OK cards still apply, a small AI model checks the last page, and when a step fails the AI picks up the job and finishes it, where a script would just stop.

When shouldn't I use it?

When the job involves choosing. If a routine should pick the cheapest flight today, a repeat would click what was cheapest last time. Leave those thinking it through each run. Nox can't repeat runs that used email, saved logins, connected apps or files in any case.

Where can I try it?

It's part of nox.cool's early access, which is by invite for now.