An AI agent that keeps going until the job is done

Key points

• Go all the way turns a task into a mission: one card with what done means, a plan, a budget ($1 to $500), a deadline (up to 30 days) and what Nox may do without asking.
• Nox works session after session in the same task. Each session starts on its own when the last one ends, or at a time Nox chose to wait for.
• What only you can do goes on one list on the mission card, with one notice; Nox carries on with everything else meanwhile.
• Before a mission counts as done, a separate reviewer checks the result against the done-checks, and Nox closes the gaps it finds.
• Paying, agreeing to terms, sending as you, posting publicly, deleting and sharing your details always ask, whatever the mission. Your own rules still win.

Most AI agents stop too soon. You ask for something real ("find me ten customers", "source a supplier", "organise the launch") and get one search, a draft, and a question back. Or they go the other way and ask permission for every click, so you end up doing the work of supervising them. nox.cool's Go all the way mode is the third way. You agree the goal, the budget and the rules once, and Nox keeps working, session after session, until a separate reviewer agrees the job is done.

#The problem: agents that give up, or never stop asking

An AI agent can browse, write, code and send. What it usually can't do is keep at it. A single sitting has a natural end: the model answers, the turn is over, and the work waits for you to say "carry on". Jobs that matter rarely fit in one sitting. They need research, then a plan, then outreach, then waiting for replies, then following up.

Two easy fixes both go wrong. Letting the agent loose with your accounts is how money and reputations get spent by accident. Asking at every step is safe but turns you into the agent's babysitter. We wanted Nox to behave like a capable person you hand a job to: agree what success looks like and what they may do on their own, then let them get on with it and come to you only when they must.

#What is Go all the way?

Go all the way turns a task into a mission: a goal Nox keeps working on in the same task until it's done. Every mission starts with one card you approve. The card holds:

• Done when: checks anyone could tick, such as "10 boxes ordered and paid" or "a spreadsheet of 20 venues with prices and contacts". These are what the reviewer judges against later.
• The plan: the steps Nox sees now. It changes as Nox learns.
• A budget for the whole mission, from $1 to $500. It covers Nox's own work: its thinking, its computer time and its searches.
• A deadline, up to 30 days away.
• What it may do without asking, chosen from four permissions, and only those the job needs.

#How do I start a mission?

1. Switch on Go all the way under the message box.
2. Describe the goal in your own words and send it.
3. Nox may look a few things up to plan well, then shows you the mission card.
4. Read it, and press Approve, or decline and say what to change (the budget, the deadline, a permission).

You can also just ask Nox to "go all the way" on something, and it will offer a mission when a job is clearly too big for one sitting.

#How does Nox keep going?

When a session ends, the next one starts on its own about twenty seconds later, in the same task, with the mission as it stands: the checks, the plan and what's done, what's left of the budget, the deadline, and anything waiting on you. Nox reads it, picks the next steps and works.

Sometimes the right move is to wait: for a supplier's reply, a delivery, a date. Nox can say "carry on tomorrow at 9" and the next session starts then. If you write in the task meanwhile, it reads your message as usual.

A mission ends when the reviewer agrees it's done, when the budget is spent, when the deadline passes, or when you press Stop. If three sessions in a row move nothing forward, it pauses as stuck and tells you, so it never spends its budget going in circles.

#What can Nox do without asking in a mission?

Only what's on the card you approved. Everything else asks as usual.

Research, installing software and writing and running code on its own computer never needed asking, because they happen on Nox's computer, not in your accounts.

Two more safeguards sit under every permission. A website step on a page about money (a checkout, a price, a card) always asks, even with the website permission. And every step a permission allows still passes nox.cool's second opinion, an independent check that the action fits what you asked for. Your own rules still win too: if you once said "never on this site", a mission can't change that.

#How does Nox fill its own gaps?

In a mission, a missing piece is a problem to solve, not a reason to stop. Nox works down a ladder:

1. Its own tools: the web, its computer, files, code.
2. Your connected apps.
3. The app catalog. Connecting an account is your click, so that goes on your list.
4. Installing software on its computer.
5. Writing its own script or small tool.
6. When something worked and you'll want it again, proposing to save it as a skill. That asks you.

And one failed try is never the end: if a source is thin, it tries another; if a way is blocked, it finds another way. It never tries to get round a refusal or a limit.

#What happens when Nox needs me?

Some things only you can do: a decision that's yours, signing in, a payment, connecting an account. Instead of stopping at each one, Nox adds them to one list on the mission card and sends you a single notice. Then it carries on with everything else.

When there's truly nothing else to do, it waits for you. Reply in the task ("done, the shop's connected", or "go ahead") and the mission carries on.

#Who checks that the work is really done?

Nox doesn't mark its own homework. When it believes every check is met, it hands its summary and evidence to a reviewer: a separate model that reads the mission's checks and the task's own record of what was done (each step and what it returned) and decides check by check. If something is missing or unproven, Nox gets the gaps and keeps going. After three reviews with gaps, the mission is handed over with the gaps named plainly, so you're never told "done" when it isn't.

#The approach round, and missions in projects

Before a big choice, a mission can run an approach round: three competing approaches (the fastest, the cheapest and the boldest), then a critic that compares their strengths and weaknesses and recommends one. Nox picks, says why, and updates the plan. A choice that's yours to make, such as money or your name on something, goes on your list instead.

In a project, the round uses your team. Each teammate proposes from their own role in the project's group chat, under their own name: marketing suggests a newsletter push, sales a wholesale trial with local cafés, product a launch price. The lead then posts the comparison. In our test, a round with three teammates cost 4 cents.

#A real mission, start to finish

Here's one we ran on nox.cool's own servers, with real web searches. The ask: "Find 6 independent bookshops in London that host author events, put them in a spreadsheet, and email each one asking whether they'd host a reading for a debut novelist in December. Before any email goes out, I want to confirm the list." For the test, the emails went to test addresses, and nothing left nox.cool.

1. Nox proposed the mission with five checks, a $6 budget and one permission, email from its own address. We approved it.
2. It ran an approach round, then researched and built the spreadsheet.
3. It put "Confirm the list of 6 bookshops, or say which to swap" on our list, and waited without sending anything.
4. We replied "go ahead". Nox sent six emails from its own address, each allowed by the mission's email permission, with no card.
5. It waited two minutes for replies, as asked; the next session started on its own and checked the inbox.
6. The reviewer agreed every check was met. The mission ended done, having spent $1.54 of its $6 budget.

#What keeps you safe

• One clear agreement. Everything a mission may do on its own is on the card you approved; nothing else changes.
• A budget and a deadline on every mission, with a live meter. Add more or stop at any time.
• Money always asks. A mission's budget is for Nox's own work. Paying for anything asks you, and Touch ID or Face ID still applies if you've switched it on.
• Your rules win. Your "never" and "always ask" rules hold inside a mission.
• An independent check on every step a permission allows, and a reviewer before "done".
• No spinning. Three idle sessions in a row pause the mission, and you hear about it.

#The short version

Go all the way is available today in nox.cool's early access. Switch it on under the message box, or ask Nox to go all the way on the next big thing you'd rather not do yourself. Sign in to nox.cool to try it.

#Questions and answers

What is Go all the way on nox.cool?

It's a mode for jobs too big for one sitting. You approve one mission card with the goal, what done means, a plan, a budget and a deadline, and Nox keeps working on it in the same task, session after session, until a reviewer agrees it's done, the budget or deadline is reached, or you stop it.

How much does a mission cost?

You set the budget on the mission card, from $1 to $500. It covers Nox's own work: its thinking, its computer time and its searches. The card shows what's been spent, and you can add more or stop at any time. Nox can't spend your money as part of the budget: paying for anything always asks.

What can Nox do without asking during a mission?

Only the permissions on the card you approved: sending forms on websites such as sign-ups and enquiries, emailing from its own address, everyday changes in your connected apps, and adding events to your calendar. Research, installing software and running code on its own computer never needed asking. Everything else asks as usual.

Can a mission pay for things or post as me?

No. Paying or booking, agreeing to terms, sending as you or posting publicly, deleting things and sharing your personal or financial details always ask, whatever the mission. A website permission never covers a page about money either.

What happens if Nox gets stuck?

It tries another source or another way first. If three sessions in a row move nothing forward, the mission pauses as stuck and you get a notice, so it never burns its budget going in circles. You can tell it what to try and press Carry on.

Does it work in a project with several agents?

Yes. In a project, the approach round asks each teammate to propose a way to reach the goal from their own role, posted in the project's group chat under their names, and the lead compares them before picking one.