How Profiles, Goals, And Agents Work
Profiles, goals, and agent runs describe different parts of long-running work.
- An Agent Profile is a reusable operating role: instructions, model policy, Skills, tools, delegation limits, and approval policy.
- A goal is the durable objective and finish condition for a piece of work that may span several turns.
- An agent run is one execution attempt with a prompt, model, status, result, and evidence.
- A parent agent coordinates the objective and combines results.
- A child agent receives a bounded subtask and returns a hand-off to its parent.
Why Delegate
Delegation is useful when independent research, inspection, or implementation can happen in parallel or needs a specialist profile. It is less useful when agents would edit the same small resource or when only one decision is needed. The parent remains responsible for deciding whether the combined evidence actually completes the goal.
Child agents do not become invisible background permissions. Their runs remain linked to the parent, and a returned result is evidence rather than automatic authorisation to write, send, publish, or declare the goal complete. Tool availability and approvals still apply to the run doing the action.
Delegation capacity is bounded. Each new run snapshots the application-wide work-round, nesting, per-parent concurrency, app-wide concurrency, and optional child active-time settings. When capacity is full, eligible children wait in a first-in, first-out queue rather than bypassing the limits.
Status And Hand-Offs
An active goal can accumulate progress across turns. Completion should mean the finish condition and required checks are satisfied. A block should identify a real impasse, not merely unfinished work. Stopping an agent ends that run without deleting the parent thread, goal history, or already-recorded evidence.
The work-round budget is checked at model, tool, and resume boundaries so a restart cannot reset it. Repeated no-progress states are blocked and then terminated cleanly, and every terminal path writes one durable final status and reason.
A useful child-agent hand-off says what was checked, what changed, what evidence was found, what remains uncertain, and whether any consequential action still needs the parent or user.
Configure And Use It
- Create and select roles with Profiles, Goals, And Agents.
- Choose a profile in a workflow with Workflows.
- Control the models behind runs in Models And Providers.
- Review waiting work and failures in Mobile And Native Surfaces or Monitor.