Blog

Do AI agents work while you're offline? Async agent tasks

What keeps running when you close the app: scheduled tasks, monitors and background work, which machine does it, and what happens when the agent needs you.

The Nebula team13 min read

Mostly yes, with one condition: the work has to be running somewhere other than the window you closed. In Nebula a task can run now, on a schedule or when something happens, on a cloud device that stays up or on your own computer if you allowed it, and it reports back in the thread that started it. Claude lists “hand off and schedule tasks” from Pro up. Slack automates through Workflow Builder.

“An agent that works while you are offline” sounds like one feature. It is four things that have to be true at once: something has to hold the queue, something has to run the commands, something has to decide what happens when the agent hits a question or a password at 2 a.m., and something has to tell you it finished. This page takes those four apart and compares what Slack, Claude and Nebula publish about each.

What “offline” actually means for an agent

People ask this with four different things in mind.

  1. You closed the app, but your computer is on. Does the agent keep running, or does it stop when the process you quit goes away?
  2. Your computer is asleep or shut. Is there a machine somewhere else that picks the work up?
  3. You are asleep. Is the work allowed to start without a person typing the request?
  4. You are out of the loop entirely. When it needs a decision, a password or a sign-in code, what happens?

A product can pass 1 and fail 2. Nebula’s documentation is explicit about the split:

  • With computer use set to Allow using my computer, agents run in your folder “even when Nebula is closed”.
  • With Allow only while Nebula is open, “the helper stops when you quit Nebula”.
  • Separately, a cloud device runs in the cloud for your workspace, and “a subscribed workspace starts with one small cloud device, which comes with the seat”.

So the answer to “does it work while I am offline” is: on a cloud device yes, on your own computer only as far as you allowed it, and in both cases only for work the agent is allowed to start. In an agent-native workspace that is the point of putting the agent in the channel rather than in your window: the channel is still there when you are not.

What keeps going when nobody is watching

Slack Claude (Pro and up) Nebula
Can work start without you? Yes, through Workflow Builder automation and “AI workflow generation”, which builds workflows from a single prompt Yes: “hand off and schedule tasks” is listed on Pro, Max and Ultra and not on Free; Team and Enterprise inherit the paid feature set Yes: a task “can run now, on a schedule, or when something happens”, and a Monitor watches for a change
Does the AI execute work or only summarise it? Slackbot is described as “your personal AI agent that understands conversations, files, and projects in Slack. Plan limits apply.” Agentic work is sold as separate surfaces: Slack Code, Agentforce, Agentic Platform Executes: Claude Code and Claude Cowork sit on the paid plans Executes: a task runs as its own thread with steps, tools and files, and reports back in the channel
Where the work runs when you close the laptop Inside Slack Anthropic’s side A cloud device, or your own computer if you allowed it
How it reaches you In Slack In Claude Email, mobile push, or a Slack, Discord or Telegram DM, chosen per event, with quiet hours and a pause

Seat prices are not the subject of this page. Nebula’s are on nebula.gg/pricing, and the side-by-side comparisons live on /compare.

Slack is the better fit for a team that already works there and wants automation beside the conversation, with SAML SSO, SCIM and an app directory Nebula does not match. Claude is the better fit for one person handing a long task to a model they already pay for. Nebula’s case is narrower: a team whose agent needs its own machine, its own memory and a channel to report into.

The four parts, one at a time

1. Something has to hold the queue

A closed app cannot hold a queue. What holds it is a task list on a server. In Nebula, scheduled tasks “are what make tasks”: a scheduled task and a monitor are both created from New scheduled task in Channel settings → Scheduled tasks, or by asking Nebula in the channel, and every run lands in one task you can open from your Inbox. Each task shows one of seven states: Queued, Running, Paused, Needs your attention, Completed, Failed and Called off. The one that matters for offline work is Needs your attention: that is where an absent person becomes the bottleneck.

2. Something has to run the commands

A cloud device is a machine your workspace owns. Sizes are published: Small is 2 CPU, 8 GB RAM and 40 GB of disk; Medium is 4 CPU, 16 GB and 250 GB; Large is 8 CPU, 32 GB and 500 GB. You upgrade one size at a time, and an upgrade is one-way. A new cloud device starts Stopped and boots the first time an agent uses it, so that first task takes a little longer. It draws Nebula Credits while it works; your own computer costs nothing.

Your own computer is the other half. It reads Online while Nebula’s background helper is running on it and Offline since 2:14 PM when it has gone quiet, and device pickers add that state after the name of a computer that isn’t online, so you can see which machines can take work right now. Work you ask for on an offline computer is refused before any task exists. A scheduled job is not checked that way yet: its run still starts and then fails on the offline computer. A cloud device is never refused, because it starts when the work arrives.

If you want scheduled work to run on your own machine without stopping to approve each command, the documentation names the switch: Settings → Workspace → Devices, open the computer, Settings tab, Let scheduled work run on this computer under Scheduled work. It is off for a new computer and only the owner can change it. And “Always allow on a card in a chat you’re watching doesn’t carry over to this either”. The three ways to give an agent a computer, and who gets to start work on each, are covered in computer use for AI agents.

3. Something has to decide what happens when the agent needs you

Every agent that touches real systems eventually hits something only a person can supply: a decision between two files, a password, a sign-in code. Nebula’s documentation answers this differently for watched and unwatched work.

A task you are watching stops and asks rather than guessing. It posts up to three questions with options, switches to Needs your attention, and the thread row reads Needs you until you answer. A task nobody is watching does the opposite: “a scheduled run or a monitor makes its best judgment and carries on, because there’s no one there to answer”.

There are two things an unwatched run still stops for. On your own computer, work nobody started asks the owner before every command unless Let scheduled work run on this computer is on. The card posts in the channel as soon as it asks, only the computer’s owner can allow it, and a step nobody answers within two days does not run. Cloud devices never ask. Separately, a write that looks risky or does not match the request is checked whatever your safety gate says; in work nobody is watching it is refused and the run reports it.

Passwords and sign-in codes have their own card. When an agent in a cloud device’s terminal hits one it cannot enter, the task shows The terminal needs a password or code and you type it into the device’s terminal yourself; the agent waits while you are there, and the password does not show on screen or in the task. The documentation describes that card for a task you are there to answer. It does not say a scheduled run will wait for one, so the practical rule is to give an overnight job the access it needs before it runs, and to write the prompt so the agent’s own judgment at a fork is one you would accept by morning.

4. Something has to tell you it finished

Nebula’s answer is a set of destinations rather than one: email is always available, and mobile push, Slack DM, Discord DM and Telegram DM appear as Connected once set up. On top of that sit rules: per-channel levels of Mentions & replies or Nothing (mute), a mute for @channel, @here and @everyone, a pause of 30 minutes up to Until tomorrow morning (Until tomorrow on mobile), and quiet hours on a weekly schedule in your own timezone that follow the person rather than the phone. A scheduled job firing at 03:00 your time should not be a phone buzzing at 03:00 your time; it should be a line waiting in your Inbox, which has a Scheduled runs tab for exactly these.

The task report itself lands where the work was asked for: the thread that started it, with the result as the body of the message and the files it made on the task’s card.

Worked example: what an overnight job actually returns

A team wants a competitor pricing table refreshed every Monday at 06:00 and posted before the standup. Mapped onto the four parts:

Part What has to be true Where it is decided
The queue A scheduled task exists on the channel and is set for Monday 06:00 Channel settings → Scheduled tasks → New scheduled task, or ask Nebula in the channel
The machine A cloud device is in the channel, or your own computer is awake with scheduled work allowed Devices settings, Let scheduled work run on this computer
The decision The run makes its own call at a fork and does not stop to ask; a risky write is refused and reported Task behaviour for work nobody is watching
The handoff Each run’s full report lands in the scheduled task’s own thread, which you open from Inbox → Scheduled runs; the channel gets one line when there’s news, and a failure @mentions the owner Notification destinations and rules
Continuity Next Monday’s run can compare against this Monday’s saved table Agent Storage, “state between scheduled runs”

The fifth row is what turns a scheduled script into an agent. A cron job that starts from zero every week cannot tell you what changed. Agent Storage is “a persistent set of files that survives across runs, threads and devices”, it “belongs to the agent, not to a channel or a device”, and the documented example is “the last invoice number a weekly job processed”. Coding agents such as Claude Code follow the same device rules when they run in your team’s channels, and stricter ones on your own computer: with nobody there they are refused rather than asking, and a scheduled job can start one there only from your Nebula DM.

Common questions

Do AI agents work while you are offline? Yes, when the work runs somewhere other than your closed window. Nebula tasks can run on a schedule or when something happens, and a cloud device keeps working whether or not your laptop is open. On your own computer it depends on a setting: Allow using my computer keeps running when Nebula is closed, Allow only while Nebula is open stops with the app.

Can a scheduled AI task run overnight without me approving every step? In Nebula, on a cloud device, yes: cloud devices never ask. On your own computer, work nobody started asks before every command unless you turn on Let scheduled work run on this computer, which is off for a new computer and only you can turn on.

What happens when the agent needs a password at 2 a.m.? A task nobody is watching never asks a question: a scheduled run or monitor makes its best judgment and carries on. It still waits for permission to run on your own computer unless scheduled work is allowed there, and a write that looks risky is refused and reported rather than run. A password or sign-in code the agent cannot enter is handled through a card that waits for you to type it into the device’s terminal, so give an overnight job the access it needs before it runs.

Will my scheduled job run if my laptop is asleep? Not on a laptop that is asleep. Nebula shows a device as Online while its background helper runs and Offline since a time when it has gone quiet, and a scheduled run on an offline computer starts and then fails. A cloud device avoids the problem because it is not your laptop.

Does the agent remember last week’s run? In Nebula, yes. Agent Storage is a persistent set of files per agent that survives across runs, threads and devices, documented for keeping state between scheduled runs so a weekly job can compare against the last one.

Is an async agent the same as an agent that runs on a schedule? No. Scheduling is one trigger; async work also needs a machine that stays up, a defined behaviour when the agent hits a question, and a way to reach you on completion. A scheduled job with none of those is a cron entry.

Sources (7)