subagent-driven-development

To prevent context pollution during implementation plan execution, dispatch a fresh subagent for each independent task to maintain isolated context and focused execution. This approach keeps each task's implementation, testing, and committing separate and clean.

Subagent-driven development works by iteratively dispatching dedicated implementer subagents for each task, addressing their questions, and running two-stage spec compliance and code quality reviews before marking tasks complete for safe delivery.

The two-stage review process enforces a spec compliance review followed by a code quality review for each task. Both stages use dedicated reviewer subagents and return explicit statuses like DONE, DONE_WITH_CONCERNS, NEEDS_CONTEXT, or BLOCKED.

This approach applies during same-session plan execution when tasks are mostly independent. If tasks require per-task implementation, testing, committing, and reviews, the controller can re-dispatch subagents on failures and manage questions.

To automate git commits during task execution, the controller subagent respects auto_commit settings while dispatching implementers. It extracts full task text into a TodoWrite and manages the commit workflow after reviews pass.

You should avoid dispatching fresh subagents per task when your implementation plan contains highly dependent tasks. This method requires mostly independent tasks to maintain isolated context and effectively run two-stage reviews without blocking.