Delegation
Delegation — Agent Orchestration & Parallelization
What It Does
Parallelizes work across six patterns: built-in agents, worktree-isolated agents, background agents, custom agents via inline briefs, agent teams via TeamCreate, and parallel task dispatch. It also splits delegation into two weights — lightweight one-shot workers (capped turns) versus full agents that iterate with tools. The Algorithm auto-invokes it once work hits three or more independent workstreams.
The Problem
Doing independent work serially wastes time, but throwing every task at a heavyweight agent wastes more — spawning a full agent for a one-shot classification burns 10-30 seconds of startup for nothing. Worse, people conflate two different systems: fire-and-forget custom agents with no shared state, and persistent agent teams that message each other and share a task list. Pick the wrong one and you either over-coordinate simple work or under-coordinate complex work. This skill routes each job to the right pattern and the right weight.
How It Works
Auto-invoked by the Algorithm when work can be parallelized or requires agent specialization.