Route and parallelize architecture work

Route and parallelize work across specialized agents

Large codebases with 5+ file changes spanning multiple domains—when parallel architecture design and implementation can run unblocked.

• · task that can be split into 3+ independent workstreams
• · effort level (Extended+) and specialization needs
• · list of subagents to spawn or team composition

• · Task() calls with subagent_type and isolation mode
• · TeamCreate() calls for persistent agent teams
• · Task coordination strategy (foreground vs. background)

• · Task tool (subagent spawning)
• · TeamCreate tool (persistent teams)
• · Git worktree (for isolation)

• · Work is parallelizable (no hard dependencies)
• · Multiple specialized subagents available

• · Confusing custom agents vs. agent teams → wrong tool chosen
• · Spawning agents without isolation → file conflicts
• · Foreground agents used when background would work → blocks main flow
• · Over-parallelizing small tasks → overhead exceeds benefit

• · Two completely different systems documented (custom agents vs. team)
• · Routing decision table for when each system applies
• · Worktree isolation with auto-cleanup logic
• · Observer team archetype documented for auditability