Everyday analogy
Before: often only one crew could work; you waited for them to finish before sending the next message.
Now: heavy work can stay with background crews while you keep talking in the main chat.
What the system actually does
It may feel like “just more parallel sends,” but under the hood:- Splits work: the main agent can turn complex asks into background subtasks
- Crews run alone: each subtask has its own progress and does not freeze the main chat
- Reports back: when a subtask finishes, it reports to the mainline and merges results into the project
- Chat stays open: while subtasks run, you can still ask new things; you may see “still generating in the background—you can keep talking”
- Progress is visible: e.g. “N subtasks running,” subtask tags in task status
- Spend is clear: subtask consumption can be reviewed
- Collaboration stays separate: tasks are tied to who started them; others generating show as “collaborating,” reducing accidental stops
- Queues are explained: when busy, items queue; when yours runs, you may see that prior work is still in the background and the queue is next
Main task vs subtask
Remember:
- Main task = commander + main conversation
- Subtask = parallel background crew
When something becomes a subtask
When the ask is large, separable, and better done slowly in the background, the main agent may spawn subtasks so you can keep asking other things (as shown in product). Flow in plain words:- You send a message (or the agent judges a piece fits background work)
- The main agent starts a subtask
- The subtask advances on its own; you can keep chatting / asking in the main thread
- Subtask finishes → reports to the main chat → results merge into the project when possible
- You can see progress in chat / task status, and subtask spend in billing
What if several people send similar asks?
In shared projects, the system tries to bill and track by who started the work: Separate by personA’s main/subtasks stay with A; B’s with B. Progress, stop, and spend stay per person so you do not stop the wrong run. Someone else generating → “Collaborating”
You can see a teammate’s work is still running. Similar ≠ auto-merged
Two people each say “make the button red” are usually two independent requests—not secretly one job. May use resources twice
Duplicate similar work can cost twice. Agree who changes what before both edit the same spot. Your own back-to-back asks
- Subtasks in background, mainline free → often you can keep asking (with a parallel hint)
- Mainline still busy / queued → you may enter the queue; when it is your turn, you get an explanation
Big picture
Value in one line
Before: one generating reply often meant waiting. Now: heavy work runs in the background while you send the next ask and watch progress and spend—like a team that can assign work, not a counter that takes one ticket at a time.FAQ
Do I create subtasks myself?
Do I create subtasks myself?
Usually no. The main agent often dispatches them when background work fits. You keep sending normal asks in the main input.
If the main task ends, is everything done?
If the main task ends, is everything done?
Not always. Ending the main task does not mean every subtask finished or every report merged.
Can I keep talking while a subtask runs?
Can I keep talking while a subtask runs?
Often yes. When the mainline is free you can still ask; the product may note background generation. If the mainline is busy or queued, you get an explanation when your turn starts.
In collaboration, could I stop someone else’s task?
In collaboration, could I stop someone else’s task?
The system tries to separate progress, stop, and spend by who started the work. Others generating show as “Collaborating,” reducing accidental stops.
Does refreshing drop subtasks?
Does refreshing drop subtasks?
Usually no. Work continues in the cloud. While parallel subtasks run, status may not be fully written back yet—refresh helps you see the latest. See FAQ.
Related docs
- Manual edit (edit elements in preview—not the same as AI-dispatched background subtasks)
- FAQ
- Changelog · 08/28

