What most AI agents still can't do: five things to check

Key points

• The basics are now common: a computer in the cloud, a browser, hidden passwords, and asking before it acts.
• Before it presses Buy, most agents show a line of text. Nox shows a picture of the real page with the button circled.
• Most products are one agent. On nox.cool you can have up to fifty, each with its own computer, inbox and memory.
• When an agent reads your email, the sign-in codes usually come with it. Nox takes them out first.
• Most agents charge the full AI cost on every run. A browser routine you've kept on Nox repeats on almost no AI.
• Suppliers can quote on a private page without ever seeing your budget.
• None of this is unique to Nox. We say "most" because at least one other agent comes close on several of them.

Most AI agents can now do the same basic things, so what sets them apart has moved. It's no longer whether an agent can open a browser. It's what it shows you before it acts, how many of them you can run, who reads your sign-in codes, and what a job costs the second time.

We checked this for ourselves, because we wanted to say on our home page what Nox does that others don't. We read the published documents of the agents we follow, then checked each claim against our own code. The result was humbling, and useful.

We looked for things only Nox can do. We found none. We found five that most agents don't.

Here are the five, and how to check them on any agent you're considering, including ours.

#Why do AI agents all sound the same now?

Because the good ones really have converged. Not long ago, an agent that could use a website on its own was remarkable. Now it's the entry fee. Every launch page promises the same thing: give it a task, it works while you sleep, it asks before anything important.

That makes choosing hard. If you're not technical, you can't tell from the words which one is careful with your money and which one just says so. The way through is to stop reading the promises and look at specific moments: the moment before it pays, the moment it reads your email, the moment it runs a job for the second time.

#What should any good AI agent do by now?

Before the five, here's the floor. If an agent can't do these, you can stop there:

• Work on a computer of its own, so it isn't taking over yours.
• Keep going after you leave, and tell you when it's done.
• Keep your passwords and card numbers away from the AI. The computer fills them in and the AI never sees them.
• Ask before anything important: sending, paying, deleting or posting in your name.
• Run a second check on risky steps, against what you actually asked.
• Let you take your things with you if you leave.

Nox does all of these, and they're not on our list of reasons to pick it, because others do them too. Passwords kept from the AI, for instance, are now common among the serious agents.

#1. What do you see before it presses Buy?

This is the moment that matters most, and it's where most agents still fall short. They stop and show you a line of text: what it's about to do, and where. That line was written by the system you're trying to check. If it misread the page, the summary carries the same mistake, and you approve the mistake.

On nox.cool, when an agent is about to press a button on a website, you get a picture of the real page with that button circled and the price read off the page itself.

Not everything gets a picture. An email shows as the email, and that text is the AI's draft, which you can edit before it goes. Plenty of smaller actions still get a plain card. But where money meets a website, you're looking at the thing.

How to check any agent: give it a purchase and stop at the last step. Ask yourself whether you could spot a wrong price from what it shows you.

#2. Is it one agent, or a team?

Most products give you one agent. It might be very good, but it's one worker with one memory, and everything you give it goes into the same pile.

On nox.cool you can have up to fifty, each with a face, a goal, and its own computer, inbox and memory. Your bookkeeper never mixes its notes with your travel planner's. You can put several in a group chat, or give a project a team and let a lead hand out the parts.

To be precise: fifty is how many you can have, not how many run at the same second. An agent that isn't working goes to sleep and keeps its files. And each one's inbox receives email for now.

How to check any agent: ask whether a second agent gets its own computer and memory, or whether it's a new name on the same one.

#3. Who reads your sign-in codes?

When an agent reads your email, it usually reads all of it, and your inbox is full of one-time codes and "click here to sign in" links. Those are keys. Text in an email can also carry instructions aimed at the AI, so the less sensitive material the AI holds, the better.

Nox takes sign-in codes and one-time links out before the AI reads the message. If a job needs one, it asks you, and the computer uses it without the AI ever reading it. This covers the email it reads to do a job, your morning look, and meeting briefs.

How to check any agent: search its documents for what happens to the codes in your email. Often they don't say.

#4. What does it cost the second time?

Most agents charge the full AI cost every single time, even for a job they did yesterday in exactly the same way. If you check the same page every morning, you pay for the agent to rediscover the page every morning.

On Nox, a browser routine you've chosen to keep can repeat its steps on almost no AI. A small one still reads the last page, so it can tell you how it went and notice if something's off. If a step doesn't fit any more, the full AI takes over rather than guessing.

On a simple page we measured a run falling from about 10 cents to well under a cent. A heavier page costs more. It only works in the browser, and you switch it on routine by routine.

How to check any agent: run the same job twice and compare the bill.

#5. Can it get you quotes without giving away your budget?

This one is narrower, but if you run a business or renovate a kitchen, it's money. Most agents leave the calling around to you, or at best write the emails.

Nox can send your request to suppliers, who each answer on a private page with no account needed. It lines up what comes back, flags what's late, missing or over budget, and asks you before choosing. Your budget stays private. The suppliers never see it, so nobody prices up to it. The AI does see it, because it has to judge the answers.

How to check any agent: ask it to get three quotes, and see who ends up knowing your limit.

#How does Nox compare?

We compared what the agents we track publish about themselves, in November 2026. These products change every week, so treat this as a snapshot. It's also why we wrote "most": on several of these rows, at least one other agent comes close.

#Where does Nox fall short?

If we're asking you to check everyone's claims, here are the limits on ours:

• It's invite-only for now. You can't just sign up today.
• There's no phone app in the stores yet. On a phone, it runs in the browser.
• Agents' inboxes receive email but don't send from their own address yet.
• Not every action gets a picture, as above.

#How do I try these on nox.cool?

1. Sign in to nox.cool. It's in early access, by invite for now.
2. For the picture: "Find me a hotel in Lisbon for next weekend, and ask me before you book."
3. For a team: add a second agent from the plus button in the sidebar and give it a goal.
4. For a repeat: set up a routine that checks a page, let it run twice, and say yes when it offers to repeat the same steps.
5. For quotes: "Get me three quotes for 500 printed boxes, and keep my budget private."

#What keeps people safe while all this runs?

• Nothing goes out without your OK, unless you've told it that it needn't ask.
• A second check reads each risky step against what you asked.
• Spending takes a held press, and can need Touch ID or Face ID.
• Everything it reads is information, never orders, so a web page can't tell it what to do.

Nox's character changed too: fewer words, no fuss, and a pair of shades when the job's done. It still doesn't joke about your money. For what else is new, see the announcement.

#Questions and answers

What can most AI agents do now?

The good ones can use a browser on a computer of their own, keep working after you leave, keep your passwords away from the AI, ask before anything important, and run a second check on risky steps.

How is Nox different from other AI agents?

Five ways: a picture of the real page before it presses a button, up to fifty agents each with its own computer, sign-in codes taken out of your email first, browser routines that repeat on almost no AI, and supplier quotes that keep your budget private.

Is Nox the only AI agent that does these things?

No. We read the published documents of the agents we track in November 2026 and found nothing only Nox does. Each of these five is something most agents don't do, and at least one other comes close on several.

Can all fifty agents work at the same moment?

No. You can have up to fifty, and each keeps its own computer, but an agent that isn't working goes to sleep with its files kept. How many run at the same moment depends on the capacity available.

Does the AI see my budget when I get quotes?

Yes. The AI needs it to judge what comes back. The suppliers never see it.

Do repeated jobs really cost nothing?

Close to it, not nothing. Only browser routines you've chosen to keep can repeat this way, and a small AI still reads the last page to tell you how it went. On a simple page we measured a run falling from about 10 cents to well under a cent.

How can I compare AI agents myself?

Give each one the same job with a real final step, and stop at that step. See what it shows you, what it would cost to run again, and what it says happens to the codes in your email.