Sub-agents handle longer, self-contained work in the background while you continue using your Mac. Each one works in its own private workspace in the cloud: it can research, read connected-app data, read specific web pages, work through large results as files, and prepare or perform multi-step work.
Delegate a task
Ask Agent Mode to delegate a clearly bounded outcome. For example:
- “Delegate a review of my open Linear issues and group them by priority.”
- “Research this company, summarize the findings, and draft a Notion note.”
- “Compare my calendar with the launch tasks and identify scheduling risks.”
Include relevant constraints, expected output, and the apps the task may use.
Follow progress
Open Agents → Sub-agents. Runs show one of these states:
- Queued or Working — the task is still progressing.
- Needs approval — a plan of changes is waiting for you.
- Needs your answer — the task asked you a question.
- Done — the result is available.
- Failed — the run could not complete.
- Stopped — you halted the run before it finished.
Select a run to open its floating panel and inspect its activity.
You can also ask by voice, in any Agent Mode conversation — including one you start later: “Did that website summary finish?” or “What did the research task from earlier find?” mrmr matches your description against your recent runs and reads back the status or result. You never need a task identifier. Ask “Show me that task” to open its panel, and “Close the sub-agent” to dismiss the panel again when you are done — the task itself keeps running.
Stop a run
While a task is queued, working, or waiting for approval, a Stop button appears at the bottom center of its panel. Click it — or press ⌘. while the panel is focused — to halt the run. The current step is allowed to finish, so the button shows Stopping… briefly before the run ends.
A stopped run keeps its activity trace, so you can still see what it got through. It does not produce a result, and any plan it was waiting on is discarded. Stopping is final for that run; delegate the task again if you want a fresh attempt.
Approve a plan
Sub-agents read freely while working, but never change anything on their own. Everything a task wants to send, create, update, or delete is gathered into one plan, and the run waits while its panel shows each change the way its app would.
Approve to make every change in the plan, or reject to make none of them. The panel then shows what each change did. A task asks twice only when one change needs the result of another, such as creating an issue and then linking it.
Answer a question
When a task cannot tell which item you meant, or something it needs is missing, it asks instead of guessing. Pick one of the choices it offers, or type your own answer, and the task carries on.
Continue the conversation
After a run finishes, send a follow-up message from its panel to refine the answer or ask a related question. The original task, its files, the result, and prior follow-ups remain available to the run. While a task is still working, a message you send steers it instead.
Tasks that ran before 7 October 2026, on the earlier background agent, stay in your history but cannot take follow-ups.
Daily limits
You can start up to 50 sub-agent tasks a day, within a daily spending limit on the work they do. When you reach either one, Agent Mode tells you which limit it was and when it resets, and nothing is started. Tasks already running finish.
When to use Agent Mode instead
Use the live Agent Mode conversation when:
- You expect quick clarification or several choices.
- You want an immediate read or a single action.
- You want to remain actively involved at every step.
Use a sub-agent when the work has a clear endpoint and may require several searches or tool calls.
