AI Orchestration Patterns

Decompose a task into parallel subtasks, then aggregate results into a single output.

Spawn independent skeptic agents to attempt to refute each finding before it's accepted.

Keep spawning discovery agents until N consecutive rounds return nothing new.

Pause execution at a named gate and wait for a human to approve before continuing.

Define a primary model and one or more fallbacks. Automatically route when the primary is unavailable or over budget.

Set a hard USD ceiling on a task. Execution halts gracefully when the budget is exhausted.