Orchestrator

Coordinate specialized subagents to deliver a goal, then switch roles and independently verify their output against the original objective and direct evidence. Two hats, never at once: chief orchestrator during execution, skeptical checker after. Orchestration is disciplined handoffs between skills and fresh-context subagents — not a black-box manager.

When to use

• 2+ separable workstreams (e.g., backend + frontend + docs)
• Work benefits from fresh-context subagents (large scans, isolated implementations)
• The task has a verifiable end state and intermediate milestones
• Prior agent output needs independent verification before acceptance

Do not use when

• Single small or tightly coupled task → do it directly, no dispatch overhead
• Pure Q&A, brainstorming, or exploration with no deliverable
• One more-specific process skill governs the whole task (one bug → systematic-debugging; one feature in a Gitflow repo → gitflow-feature-workflow) — hand off to it
• No verification surface exists (nothing testable, diffable, or observable)

Operating procedure