Platform

Background Jobs

Some work takes minutes. Some takes hours. You should not have to babysit chat while the agent grinds through fifty companies or watches a long process complete. Background Jobs is the surface where work moves out of the live conversation and into a panel you can check in on. The donut under the composer spins while there is unfinished work.

Updated today

Overview

At a glance
Indicator
Donut under the composer, spins while any job is in flight
Panel
Click the donut to open the Jobs panel
Job lifecycle
Created, running, succeeded, failed, cancelled
Live updates
Status and steps update without a refresh
Artifacts
Files, canvases, and notes the job produced appear inline
Missions
Higher-level objectives with a plan and child jobs

Most chat turns finish in seconds. Anything that does not finish in seconds is a candidate to push into the background. The platform decides automatically when to do this so you do not lose chat responsiveness. You always know there is something running because the donut spins.

The donut indicator

The small circular indicator beneath the composer.

When at least one job is running, the donut animates. The fill ring reflects approximate progress when the job exposes step counts. Hovering shows a quick label of how many jobs are active. Clicking opens the full Jobs panel.

Donut stateMeaning
Static and emptyNo work in flight
Spinning, partial fillOne or more jobs running, progress reported
Spinning, indeterminateJobs running but no step count available
Red tintAt least one job has failed and is awaiting your review
Green pulseA job has just succeeded and is awaiting acknowledgement

When work goes to the background

The platform pushes work out of the live turn for any of three reasons.

TriggerExample
The work has a long natural runtimeScraping a list of 200 prospects, building a dataset, transcribing a long recording
The work polls or waitsWatching a workflow finish, waiting for a signature on a contract, retrying a failed step
You asked for it to run laterScheduling a job to fire at a specific time or on a recurring cadence
Note
You can keep chatting while a job runs. Future messages can reference the job (for example, ask for a summary or to act on the result). The agent picks up the right context automatically.

Warm and cold work

Background work runs in one of two modes. You never have to choose — the platform routes it.

When the agent runs something in the background, it runs in one of two modes. The difference is whether the result comes back in the same turn or wakes a new one later.

WarmCold
RecordedIn memory only, no panel rowA row in the Jobs panel
ResultFolds back into the same turnWakes a new turn when it finishes
SurvivesThe current turnA closed tab, a new session, hours later
Best forWork the agent needs the result of right nowLong work you want to walk away from

A warm job is the default for a command whose output the turn depends on: the agent kicks it off, keeps working, and the chat loop folds the result back in as soon as it lands, holding the turn open just long enough for whatever was waiting on it. A cold job is for work that outlives the moment — it gets a row in the Jobs panel, survives a closed tab, and pages the agent in a fresh turn when it completes.

The two share the same machinery; the only difference at launch is whether a completion callback is attached. That makes promotion automatic: if a turn reaches its limit while warm work is still running, the platform hands that work to the cold path so the result still arrives later instead of being lost.

Watching a live process with monitor

Some work you want to watch, not just wait on. The monitor tool attaches to a long-running command and streams its output live into chat under a monitor card, tailing the log as it grows. It stays in the turn up to a cap of about fifteen minutes; if the process runs longer, it hands off to the cold path and wakes you when it finishes.

Cold-wake is verified
When a cold job finishes it calls back to wake the conversation, and that callback is signed and verified before anything acts on it. If many parallel jobs finish at once, only the last one wakes the chat — draining every settled result together — and a circuit breaker suppresses wake storms when something is retrying in a loop, with a visible notice when it does.

The Jobs panel

One panel, all running and recent jobs.

Click the donut to open the panel. Each row is one job. The header shows the title, the owning agent, the status, and the elapsed time. Expanding a row shows the plan (the named steps), the live status per step, and any artifacts. Jobs from the last seven days stay visible by default; older ones are searchable by name.

ColumnWhat it shows
TitleShort name the agent gave the job
AgentWhich agent owns this job
StatusRunning, succeeded, failed, cancelled, or paused
StepsCurrent step number out of total, plus the step kind
StartedWhen it kicked off
DurationTime elapsed for in-flight jobs, total time for finished ones

Missions

When the work is bigger than one job, it becomes a mission.

A mission is an outcome you want, broken down into a plan of steps. Each step is one job. Missions are how the platform handles work like building a campaign end to end, researching and then writing about a topic, or driving a multi-day deal motion. You can monitor the mission as a whole, expand it to see each step, and intervene if something looks wrong.

Step kindWhat it does
ResearchWeb searches, profile lookups, evidence gathering
SkillRuns a specific named skill an agent owns
SandboxSpins up the Computer tab and runs commands or builds artifacts
CarouselGenerates a structured visual artifact like a slide deck
SubagentHands off to another named agent for a focused subtask
HitlHuman in the loop. Pauses the mission until you confirm or edit
Mission stepGeneric step kind for plan items that do not fit the others
Tip
Missions are best for outcomes you would otherwise describe as a project. Single requests can almost always stay as inline turns. Reach for a mission when you would have written a checklist in a doc.

Monitoring a job

How to know what is happening without sitting on the page.

  1. 01
    Open the panel
    Click the donut. The panel slides in from the side.
  2. 02
    Expand the row
    The plan opens with each step in order. In-flight steps stream their progress live.
  3. 03
    Click any step
    Shows the input, the work the step did, and the output. Long outputs are scrollable in place.
  4. 04
    Leave it and come back
    Notifications surface on completion. The donut turns green on success, red on failure.

Cancelling and rerunning

Stop, restart, or pick up where a job left off.

ActionWhereWhat happens
CancelJob row menuStops the running step. Marks the job cancelled. Anything already produced stays.
Retry the failed stepJob row menu, only when status is failedReruns the failed step with the same input. Useful for transient errors.
Rerun from scratchJob row menuCreates a fresh job with the same brief. Old job stays in history for reference.
PauseJob row menu, only on long missionsHolds the next step until you resume.
ResumePaused job rowContinues from the next step.

Use cases

Five concrete examples of background work.

Build a prospect list

You ask Cortex to find 200 SaaS founders in Europe matching a profile. Cortex queues a background mission. Step one is the search, step two is the enrichment, step three is writing the list to a canvas in Builder. You keep chatting about other things while it runs. The donut turns green when the canvas is ready.

Run a daily content drop

You schedule Pulse to draft a LinkedIn post every weekday at 8am, pulling from recent activity and your brand voice. The schedule is a recurring mission. Every weekday morning a new job appears in the panel with the draft ready for your review.

Watch a long workflow finish

A carousel generation can take a few minutes for a multi-slide deck. Specter and Pulse work together. The donut spins. When the deck is ready it lands in Builder and the donut turns green.

Drive a deal end to end

Striker takes a deal from first reply to proposal. The mission has steps for the discovery summary, the proposal canvas, sending it for signature, and tracking the signature status. Each step appears as a job. Some steps pause the mission with a human in the loop so you can review and approve before the next step runs.

Long-running scrape with checkpoints

A multi-source data pull where each source can be slow runs as a mission with one job per source. If one source fails you can retry just that step without rerunning the others.