Subagents Studio

The main assistant coordinates work that can run separately, such as web research, coding, notes, or scheduling. Agents Studio shows those background tasks and their activity.

Why work is delegated

Delegation keeps the main conversation responsive and applies a narrower tool profile to specialized work. Current-information and coding requests can be routed automatically. The main assistant can also launch a background agent as part of a plan.

Read the task list

Each entry identifies the profile, status, task summary, start or update time, and activity. Open Task Details to inspect progress, tool events, warnings, and the final report.

Parallel tasks may finish in a different order from the order they were launched. Use task identity rather than transcript position to correlate results.

Steer a background task

Provide concise guidance that changes the existing objective without replacing it accidentally. A steer can arrive immediately or after the current tool call reaches a safe boundary.

If you need materially different work, launch a separate task so both scopes and results remain clear.

Stop a task

Stop an agent that is working on the wrong target, consuming unnecessary resources, or waiting on a condition that no longer matters. Stopping a background task does not stop the Cloud VM or delete work already written through approved tools.

Approvals from background agents

Background tools use the same central approval inbox. The card identifies which agent requested the action. Review the exact arguments and diff; do not approve merely because the delegated profile is trusted for other work.

Context and limits

Subagents receive their task, selected knowledge, and permitted tools, not the entire private history by default. Recursive spawning and task context are bounded by runtime policy. Summarize crucial constraints in the delegated task when they are not obvious from the current request.