An AI agent that does many things at once

Key points

• Each agent has one chat. Quick questions are answered there; work that takes more than a few steps becomes a job, a live card in the chat.
• Jobs run side by side. Each works in its own browser tabs on the agent's computer, so two can browse at the same moment.
• Only the whole desktop takes turns (clicking by position, recording the screen), and the person's own work goes first. In nox.cool's early access that's about 3 in every 100 steps.
• Questions from work that runs on its own wait 20 minutes, then the work carries on safely or stops and says what it needs. Nothing else is held up meanwhile.
• Jobs ask before they spend, send or sign, exactly like tasks always have.

An assistant that can only do one thing at a time isn't much of an assistant. Nox, the AI agent on nox.cool, now keeps one ongoing chat with you and does real work as jobs beside it: several at once, each in its own browser tabs, each ending with its result right in the chat. Nothing you ask waits behind something else. This post explains why we changed it, how it works and what keeps you safe.

#Why do AI agents make you wait?

Most AI agents work like a single worker at a single desk. You hand over a task, the agent opens its browser and gets going, and anything else you ask joins a queue. If the first task takes ten minutes, so does your one-line question.

That's tolerable when tasks are short. It stops being tolerable once an agent does real work: comparing flights across five sites, filling in a long form, reading a pile of invoices. And it gets worse when the agent also works on its own, on routines and check-ins you set up weeks ago.

We found this out the hard way, using Nox ourselves. One of us sent Nox a YouTube link and asked what was in the video. Nothing happened, except a line that said "Waiting its turn". The reason was four days old: a weekly check-in had asked a question on a Thursday, nobody saw it, and every new request to that agent queued up behind it. Typing into other tasks got "still working on another task".

The fix wasn't a better queue. It was no queue.

Nothing about that was really one-at-a-time work. The check-in needed an answer, not the agent's computer. The YouTube question needed a few seconds of reading. They only blocked each other because the agent was treated as busy as a whole.

#What we changed

Three ideas, each simple on its own:

#How does the chat work?

Open an agent from the side menu and you land in its chat: one conversation that keeps going, the way you'd message a colleague. The agent answers quick things itself, straight away: a sum, a fact, a short draft, a question about something it did yesterday.

When you ask for something that takes real work (anything that needs the browser, the agent's computer, research or more than a few steps), the agent starts a job and says so in a line. The job shows in the chat as a card that follows it: working, needs you, or done. The chat stays free, so you can keep talking.

When a job finishes, its card shows the result. Your next message can lean on it without repeating anything: "which one was cheapest?", "book the 9:40", "send that to Sam". The agent is told when a job ends, so it knows what you're referring to.

You stay in charge of jobs from the chat:

• "What's running?" lists them.
• "Tell the flights job to look at Porto too" passes a message to a job while it works.
• "Stop the flights job" stops it.
• Tap a card to open the job and watch it work, step by step. Each job is also a task in your list.

The + next to an agent still starts a separate task directly, if you'd rather.

#How can one agent do several things at once?

Each agent on nox.cool has its own cloud computer with a real browser. The insight that makes parallel work possible: almost everything an agent does happens inside a browser tab, and clicking, typing and reading a page all work in a tab that isn't in front. Two tasks can share one browser as easily as you can have two tabs open.

So each task now works in tabs of its own. The first time a task browses, it gets a tab; links it opens become its tabs too. A task sees only its own tabs and can't touch another task's. When it finishes, its tabs are free for the next one.

What can't be shared is the screen itself. A few steps need the whole desktop: clicking at a position on the screen (for apps that aren't web pages), or recording the screen. Those take turns, and if your own work and a background routine both want the desktop, yours goes first. A task waiting for it says "Waiting for the desktop" and names what's using it.

How often does that happen? Over the last two months of nox.cool's early access, about 3 in every 100 steps on agents' computers needed the whole desktop. The other 97 happened in a tab, and now run side by side.

#What happens to a question nobody answers?

Your own tasks still wait for your answer as long as you like. Questions show in the task and in your Inbox, and typing into a task that asked you something now counts as the answer.

Work that runs on its own is different: routines, triggers, check-ins, long missions. When one of those asks a question and nobody answers within 20 minutes, it carries on with the safest reasonable choice and says so in its report, or it stops and tells you what it needs. The same 20 minutes already applied to its requests for an OK. And since an agent is never busy as a whole any more, a waiting question holds up nothing else in the meantime.

#What does this look like day to day?

A few real ways people use it:

• "Find me flights to Lisbon in March under £150." A job starts. Meanwhile: "what's the time there now?" answered at once.
• "Draft this week's newsletter from my notes, and find three plumbers near me with good reviews." Two jobs, side by side, each with its own card.
• "Go through my invoices folder and tell me what's overdue." A job reads the files while you ask the agent to reword an email in the chat.
• A morning routine checks your calendar and inbox in the background while you ask for a quick sum. Neither waits for the other.

#Your coding sessions, from your phone

The same idea extends to your Mac. If you use Claude Code or Codex, those sessions keep working on your computer while you're away from it. With nox.cool for Mac 0.1.9, turn on Settings › Your Mac › Coding in the Mac app. Then the Coding page on nox.cool, from any device, lists each session with where it stands. Open one to read what was asked and said, and type what to do next: it continues on your Mac.

What tools printed is never sent, only what was asked, what was said and a line per step. A session that's busy in a terminal isn't interrupted. You can also list ports (5173, say) so your agents can open the app you're building at mac.nox:5173 and check it in their browser. You can ask an agent too: "what's my VAT session up to?"

#What keeps you safe

Doing more at once doesn't loosen anything:

• Jobs ask first, exactly like tasks. Anything that spends money, sends a message, signs or posts waits on an OK card, and the job's card in the chat says it needs you.
• Your spending limits cover everything together: each job's cost shows on it, and the daily limit counts all of them.
• The chat itself never uses the computer. It answers or hands the work to a job, so a long job can never tie up your conversation.
• Tasks can't touch each other's tabs. A step aimed at another task's tab is refused.
• Background work never jumps ahead of you for the desktop.
• Questions only lapse for work running on its own, never for something you started yourself.

#Questions and answers

#Can an AI agent do several tasks at the same time?

Nox can. Each agent has its own cloud computer, and each task works in its own browser tabs on it, so several tasks browse, read and fill in pages at the same moment. Only steps that need the whole screen take turns.

#What is a job on nox.cool?

A piece of work an agent starts from its chat when something takes more than a few quick steps. It shows as a card that says whether it's working, needs you or is done, and ends with the result. Each job is also a task in your list.

#Why did my AI agent say "waiting its turn"?

It used to mean the agent was busy with another task, and once, a background check-in waiting on an unanswered question held everything up for days. Now each task is busy only with itself, and questions from work running on its own lapse after 20 minutes.

#Can I keep talking to Nox while it works?

Yes. Ask another question, start another job, tell a job something or stop one, all from the chat.

#Does running jobs at once cost more?

No more than running the same tasks one after another. Each job's cost shows on it, and your spending limits apply to all of them together. Quick answers in the chat use a lighter, cheaper brain when that's enough.

#Can I continue my Claude Code sessions from my phone?

Yes, with nox.cool for Mac 0.1.9: turn on Settings › Your Mac › Coding in the Mac app, then open Coding on nox.cool from any device.

#Try it

Sign in to nox.cool, open Nox, and ask something quick. Then ask for something big, and keep talking while it works. What else is new this week is in one chat per agent, and Nox does many things at once.