Follow one Task.Request to review.
A request becomes a Task. The Task carries its sources. An agent attempts it. A person reviews the result, and whoever holds publish decides what ships.
Six moments.Three of them are people.
Morrow, a two-person studio, needs a new headline on its spring collection page. This is everything that happened to that one Task.
Your part is four small things.
- Ask
Say what you need
In a Channel or straight in a new Task. A clear outcome and a date are enough.
- Point
Link the sources
Reference the brief and the guide from the Library, so the work starts from your words.
- Assign
Give it an owner
A colleague or an agent. Assigning a Task to an agent starts its first Run.
- Decide
Review the result
Comment, ask for changes or approve. The request and the result sit in the same thread.
The agent reads what it may.The rest is not there for it.
Shop Builder works in the Shop Space. The wholesale price list lives in Studio, where it has no access. Its search comes back empty, and the Run goes on and says what it could not find.
Sources come from the Task and the Library
A reference can point at exact lines of a page, pinned to the revision that was read, so a result can be checked against its text.
A step it may not take becomes a request
Give publish to the people who should decide. At Morrow that is Martin and Lena, so Shop Builder opens the Change Request and asks.
Three answers a person can give.
Each one is recorded on the Task with who gave it. Only one of them lets the change out.
Ask for changes
A comment in the Task thread that mentions the agent. The mention starts a new Run that picks up from the comment. The first Run stays in the thread.
A new RunApprove this version
Approval is bound to the version he previewed. If the branch moves afterwards, publishing is refused and the review starts again.
PublishedDeny or stop
Deny the request, or stop a Run while it works. Nothing is published, and the Task keeps what was done so far.
Denied
The same path works with no agent at all.
Assign the Task to a colleague and the thread, the sources and the review stay exactly where they are.
How work starts and stops.
Does every message start agent work?
No. Work starts when you assign a Task to an agent, mention it from the @ menu, or an Automation fires. In a Channel, an agent that belongs there can also pick up a message that is clearly a request for it. Editing or resending a message never starts a second Run.
Can an agent change anything it wants?
No. An agent acts with its own permissions in each Space, checked on every action. A change to those permissions applies from its next action. See roles and approvals.
What happens when a Run fails?
The Run is marked failed with the reason, such as a usage limit or an expired sign-in, and offers the matching next step. A retry is a new Run linked to the first, so the history shows both. A failed Run never posts a made-up result.
Can I correct an agent while it works?
Yes. Write to it in the thread and the instruction reaches the running attempt at its next safe point. You can also stop the Run.
Where does the agent actually run?
On a Work Machine: a Linux computer you connect, running Claude Code or Codex signed in with your own plan or API key, or on QUESTPIE Managed, which is on request during the beta. See bring your AI.
Run your first Task this way.
Free during the open beta. Bring one request and one colleague, human or agent.