Orchestration and delegation
How a Code task routes work, permissions, and optional delegated agent work while keeping decisions visible.
Route a task
The lower-left composer control is labeled Mode, permission, delegation. Its choices describe how a task may proceed: No delegation, Manual approval, or an automatic mode when the selected runtime supports it. The model chip controls the active runtime, model, and intelligence setting. These controls work together: the selected runtime determines which model, effort, permission, and delegation combinations are actually available.
Delegate selected work
When delegation is supported, a lead task can hand a bounded piece of work to an available agent. The task remains the place where the request, progress, and returned result can be reviewed. A delegated handoff can carry the selected plan, file list, or failing test context needed for the assignment. Vendor-specific memory, login sessions, and Git state do not transfer as if they were one shared runtime.
The Agents tab in Task tools shows agent activity beside the task. Settings → Agents provides the installed agent and profile controls that are available on the desktop. A manual mode keeps an approval point before delegated work begins. Automatic modes remain subject to the selected permission ceiling and runtime capability.
Set a permission boundary
The permission menu offers Read only, Project write, Ask as needed, and Full access where the runtime supports each choice. A read-only task can inspect evidence without changing project files. Project write permits edits within the project boundary. Ask as needed pauses for decisions that require approval. Full access is available only when the current policy and runtime allow it. Scheduled Code tasks use the same concepts through Permission ceiling and Delegation fields.
Antigravity and custom runtime routes do not expose delegation, and Antigravity cannot use Ask as needed. The account-backed Agents roster supplies shared profiles; Code agents use them for local assignments with the task's runtime and permissions. On mobile, the paired Remote task can display agent activity and approval requests, but the phone does not run a local Code agent. Tasks, Runtimes, Skills, and Privacy explain the surrounding boundaries.