do-delegation
Delegation: Agent Orchestration and Parallelization
What It Does
Delegates bounded work through the subagent tool your harness exposes (Agent in Claude Code and Pi; the equivalent tool elsewhere). It distinguishes direct work from delegated work, independent workers from coordinator-managed role cohorts, foreground results from background execution, and shared-checkout work from worktree-isolated edits.
Core Rule
Delegate only when specialization, context isolation, or parallel execution saves more time than dispatch and verification cost.
Independent agents do not share context, memory, task state, or results. The parent conversation remains the coordinator. Every agent prompt must be complete, and the parent must validate each result before using it.
Current Agent Contract
Use the subagent tool supplied by the active harness. Do not write source-code snippets that pretend to call the tool. If the harness has no subagent tool, do the work directly and say so.
Use only fields exposed by the current tool schema. The portable core, named as in Claude Code and Pi (map each to the active tool's equivalent field):