subagent-driven-development

Read the plan once, create a task per item, then dispatch a fresh implementer subagent per task with a scoped brief file. After each task, write the diff to a review file and dispatch a unified reviewer before marking the task complete.

Subagent-driven development runs in the same session with a fresh subagent per task and automatic review checkpoints. Executing plans runs in a parallel session with human-in-the-loop handoffs between tasks.

Always specify the model explicitly when dispatching. Use cheap models for mechanical single-file tasks with complete specs, standard models for multi-file integration, and the most capable model for architecture and final review work.

A BLOCKED status means the controller must change something before retrying: provide missing context, re-dispatch with a more capable model, split the task, or escalate plan errors to the human. Never force the same model to retry unchanged.

The .tasks.json ledger co-located with the plan records each completed task's status and commit range. On resume, reconcile it against git log and never re-dispatch tasks already marked completed.

Avoid it when tasks are tightly coupled and cannot be delegated independently, when no written implementation plan exists, or when a parallel-session workflow with human review between tasks is preferred.