Blog

What happens when an AI agent task fails?

What Failed, Needs your attention and Called off mean for a Nebula task, who gets told, who can answer or stop a run, and what pauses when credits run out.

The Nebula team14 min read

A failed agent task is rarely a dead end. In Nebula a task shows one of seven states, and only one of them is an error: Failed means something broke, Needs your attention means the task stopped on purpose to wait for you, and Called off means a person cancelled it. Most products describe agent work as either done or errored and skip the third case, which is the one that decides whether you can leave an agent alone: the task stopped because it needed a decision. In Nebula that case has its own state, its own cards, its own notification event and a rule about who may answer.

This page covers what each state means in the documentation’s own words, what stops and waits for a person, who can answer and who can stop a run, what a scheduled run does when nobody is watching, and what happens when credits run out. It is about Nebula. A short section near the end says how Claude Tag and Slack differ, and what this page does not claim about them.

Four things people mean by “fails”

“What happens when it fails” hides four separate situations. The run errored: a tool call threw, a connection died, a step timed out. The run stopped for you: it reached an action it is not allowed to take alone and is waiting. The run could not be paid for. Or a person stopped it on purpose because the work was wrong or no longer needed. Only the first is an error. The other three are product design, and Nebula gives each one its own state rather than folding them into “failed”.

The seven states

Nebula’s task documentation lists the states and defines each one. Quoted, not paraphrased:

State What it means
Queued Waiting to start
Running Working now
Paused Stopped for now, not working. Rarely seen today: a task that stops for you shows Needs your attention instead
Needs your attention Waiting on you for an approval, an answer or a connection. The task’s header and its card in the chat shorten this to Needs you
Completed Finished, with its result on the card
Failed Stopped by an error or an interruption. On desktop, its card in the chat says why when the cause is known
Called off Cancelled before it finished

Three of those rows are about work that did not finish, and none of them is the same. A task that needs you has not failed; it has stopped at a checkpoint and is holding its place. A task that was called off did not break; someone decided it should not continue. Only Failed means the machine hit something it could not get through.

The same vocabulary reaches inside the work. When a task hands pieces of itself to helper agents, each helper’s card says how it ended, such as Done, Failed, Timed out, Interrupted or Cancelled, so a run that mostly worked shows you which branch did not. And a thread that asked for a dozen tasks still costs the conversation one line, such as 4 completed · 1 failed · 1 stopped.

When a task stops for you

When a task cannot go on without you, it stops and says so. The card lands where the work is being talked about, inside the task’s thread on desktop, and the task switches to Needs your attention. The documentation names four kinds of card, and the write-approval safety gate adds a fifth:

  • A device action. When an agent wants to run a command or touch a file outside what its device allows, the card reads Computer control with the action, and shows the command and the folder. Approve or deny. Work nobody started from Nebula, such as a scheduled task, a monitor or a message that arrived from Slack or email, gets a different card that names the computer and the command, and only that computer’s owner can allow it to run (the teammate whose run asked can still deny it).
  • A plan. “Review the plan before the work starts”, with Approve plan, or Not this, I’ll reply to say what to change.
  • A question. Up to three questions, each with options and an Other box. “The task stops to ask rather than guessing, so nothing is left running while it waits.”
  • An app to connect. “Connect Gmail to continue”, with Connect or Decline. Once the app is linked, the card closes and the task starts again.
  • A risky write. With the safety gate on, a write that looks risky or does not match what you asked for pauses the task on a card reading Approve: and the action. The gate is off by default and each person turns it on for themselves.

Once answered, the card leaves a one-line result such as Approved or Connected, which is the difference between an audit trail and a transcript you have to re-read. Needs your attention, Needs you and Waiting on you all mean the same thing: the task is waiting for you.

Two edges matter before you rely on any of this. First, a request from unwatched work to run a command on someone’s own computer does not sit forever: “A step nobody answers within two days doesn’t run”. Cloud devices never ask, which is one reason work meant to run around the clock belongs on one; computer use for AI agents covers that choice. Second, a task nobody is watching never asks a question. “A scheduled run or a monitor makes its best judgment and carries on, because there’s no one there to answer”, and a write in that work that looks risky is refused rather than run.

Who can answer, and who can stop

The rules here are deliberately narrow. A risky-write card can only be answered by the person the task works for; everyone else sees the buttons greyed out. A device card for a command on someone’s own computer can only be answered by that computer’s owner. When a task leaves its browser waiting for a sign-in or a code, only the person the task works for sees the live page and its buttons, and “a browser left waiting closes on its own after 20 minutes”.

Stopping is just as bounded. You can click Stop in a task’s header, or hover the state icon on its card and choose Stop this task, and while a turn is going, the composer’s send button turns into a stop button. “Only the person who started that run can stop it - for a scheduled task or a monitor, that’s whoever created it.” Anyone else sees Stop greyed out. On a tool where anyone can cancel anyone’s run, a teammate who disagrees with a direction can quietly end the work. Here the run’s owner is the one who can call it off, and everyone else can watch.

When a scheduled run fails, or cannot be paid for

A scheduled task keeps running until you pause or delete it, or until Nebula pauses it: after ten failed runs in a row, when an app it uses needs reconnecting, or when credits run out. After each run, Nebula tells the channel when there is something worth saying. When a run fails or stops for a decision from you, it posts in the channel, mentions the task’s owner and links to the job. If it keeps failing for the same reason, it says so once rather than after every run. Routine runs do not leave status lines in the chat, on desktop or on your phone, so a channel whose jobs run hourly does not fill with run lines. To see what ran, click Nebula at the top of the channel: Recent runs lists the last jobs, newest first, with Failed on one whose last run failed.

Recovery is a person’s action. If the run stopped on a card to reconnect an app, reconnecting it starts the task again; if the work needs more, reply in the task’s thread, or use Continue this task… for a run with no conversation of its own.

The third kind of failure is the most common in practice and the least discussed: the task could not be paid for. A scheduled task “pauses itself when the workspace has no subscription or has run out of Nebula Credits”. The schedule pauses rather than failing run after run. A run that was already under way when credits ran out stops with a note saying so, and the schedule is still there when you top up: it picks up again on its own after a top-up by card, but after buying credits in the mobile app you resume it yourself. The pricing page sets the same expectation from the other side: “Nebula asks you to top up before it spends more, so nothing runs up a bill you did not expect.” Credits cover “only tasks for Nebula’s agents, and the tools and computer hours they use”; messaging and calls are included, and an agent you bring yourself runs on your own subscription at no extra charge.

Being told

Nebula’s notification rules are explicit about what travels and what does not. “A rule sends one kind of event to the destinations you pick. An event with no rule isn’t sent anywhere”, so Nebula starts you with rules for mentions, approvals, call invites, apps that need reconnecting and device problems. Task failed and Job failed are separate events: the first fires when a task runs into an error, the second when a scheduled task’s run fails. Approval needed fires when a task is waiting for your approval, and Reconnect needed when a connected app needs you to sign in again.

Email is always available as a destination. Push to your phone, and Slack, Discord or Telegram DMs, appear once connected. If you want to be told when something breaks, that is a rule you can see and edit in Settings, not a promise buried in a help article. The documentation’s own advice is to route Approval needed to your phone, so a task waiting on you reaches you wherever you are.

Worked example: the digest that never posted

Nebula’s support documentation walks through the case this page is about:

You: Nebula, my daily competitor digest didn’t post this morning. What happened?

Nebula: The 9 AM run started but stopped at the first step - your Gmail connection expired overnight, so the task couldn’t read the newsletters it summarizes. There’s a card in #market-intel to reconnect Gmail. Once you do, the job runs again on its schedule.

Read what did not happen. The run did not loop on failing auth, and it did not vanish. It stopped at a known checkpoint, left a card where the team could act on it, and said what would happen next. That is the practical test for any agent product: when it hits a wall at 2 a.m., does it leave something a person can act on in the morning?

Reading a failed run

A failure you cannot inspect is just a bad outcome. Inside a Nebula task you see each step as it happens, “including the tools it ran (click one to see its input and output), a To-do list for longer work, and how long it worked”. Helper cards end with how they finished, such as Done, Failed, Timed out or Cancelled, with a Delegating strip showing how many are finished. Web browsing streams live inside the task, with Take control to use the page yourself and Release control to hand it back. A pull request the task links shows as Open, Draft, Closed or Merged, with a tick, a cross or a dashed circle for checks that passed, failed or are still running. Scratch files an agent writes while working, “probe scripts, notes”, stay inside the task and do not post into the channel, so a failed attempt does not clutter the thread.

That is the difference between “the agent did something weird” and “step 4 stopped because the HubSpot connection needs reconnecting”.

Where Claude Tag and Slack differ

Claude Tag is Anthropic’s shared @Claude that joins your Slack channels and runs multi-stage tasks in the thread; it requires a Claude Team or Enterprise plan. Slack’s own AI is a tier of the Slack seat and works on the conversations and files inside Slack. Neither vendor’s public pages describe what a failed task looks like, who may answer or stop one, or what happens to an unanswered approval, so this page makes no claim about how either handles failure. The sourced head-to-heads are Nebula vs Claude Tag and Nebula vs Slack, with the rest on the comparison hub.

If your team already lives in Slack and wants an agent in those exact channels without moving, Claude Tag is the lower-friction path, and Slack’s bundled AI already covers “AI inside our chat tool”. This page is for the team that wants the failure path itself, who is asked, who is told and what is recorded, to be documented product behaviour, which is what an agent-native workspace is for.

Common questions

What happens when an AI agent task fails? In Nebula the task shows one of seven states. Failed means “stopped by an error or an interruption”, and on desktop the card says why when the cause is known. Needs your attention is a different state: the task stopped on purpose for an approval, an answer or a connection, and is holding its place. Called off means a person cancelled it before it finished.

What does “Needs your attention” mean? The task is waiting on you for an approval, an answer or a connection, and it is not an error. The card lands in the task’s thread, and once answered it leaves a one-line result such as Approved or Connected. Needs your attention, Needs you and Waiting on you all mean the same thing.

Who gets told when an agent task fails? A scheduled run that fails or stops for a decision posts in the channel, mentions the task’s owner and links to the job, once per cause rather than every run. Outside the app, whichever destinations your rules send Task failed and Job failed to: email is always available, and push to your phone or Slack, Discord and Telegram DMs once connected. An event with no rule is not sent anywhere.

Why did my scheduled task not run? A scheduled task pauses itself when the workspace has no subscription or has run out of Nebula Credits, and it picks up again once that is fixed: on its own after a top-up by card, or resume it yourself after buying credits in the mobile app. Routine runs do not leave status lines in the chat, so check Recent runs under Nebula at the top of the channel, which marks a job whose last run failed. If a run stopped on a card to reconnect an app, reconnecting it starts the task again.

What happens to work in progress if I stop a task? It becomes Called off, “cancelled before it finished”. While a turn is running, the composer’s send button turns into a stop button. Only the person who started the run can stop it; for a scheduled task or a monitor that is whoever created it.

How long does an unanswered approval wait? For a request from unwatched work to run a command on someone’s own computer, “a step nobody answers within two days doesn’t run”. A browser left waiting for you closes on its own after 20 minutes. In both cases the work stops, and you decide whether to start it again.

Sources (4)