subagent-driven-devlisted

# Subagent-Driven Development (SDD) Dispatch focused subagents for each task in an implementation plan, then review results in two stages. This maximizes parallelism and keeps each agent's context small and precise. ## When to Use - Implementing a plan with 3+ independent steps - Changes that span multiple packages or domains (e.g., backend + frontend + tests) - When tasks can be parallelized (no sequential dependency between them) - Do NOT use for single-file changes or tightly coupled sequential work ## Protocol ### 1. Decompose Break the work into independent, self-contained tasks. Each task must: - Have a clear, measurable "done" condition - Be completable without knowing the results of other tasks - Touch a distinct set of files (no overlapping edits) If tasks share files, they are NOT independent — sequence them instead. ### 2. Dispatch For each independent task, launch a subagent with: - **Precise instructions**: What to change, in which files, and what "done" looks like - **Context-free**: Include all necessary context in the prompt — don't assume the subagent knows what you know - **Test requirement**: Each subagent must write or update tests for its changes - **Isolation**: Use `isolation: "worktree"` for tasks that modify overlapping build artifacts Launch independent subagents in parallel (multiple Agent tool calls in one message). ### 3. Model Selection Match model to task complexity: - **Mechanical** (rename, move, update imports, config changes): `h