subagent-driven-development

Subagent-driven development coordinates multi-task plan execution by dispatching an isolated implementer subagent per task, followed by spec compliance and code quality review loops. It ensures consistent correctness and maintainability across mostly-independent development tasks within a single session.

To prevent context contamination, this approach constructs isolated prompts for each task and dispatches a fresh implementer subagent. This prevents context pollution and keeps work focused on the specific plan task without prior task interference.

Spec compliance and code quality are enforced through a two-stage review checkpoint. A spec compliance review runs first, followed by a code quality review, with iterative fixes applied until both review stages pass successfully.

Yes, structured escalation handling uses clear status outcomes including DONE, DONE_WITH_CONCERNS, NEEDS_CONTEXT, and BLOCKED. These statuses determine whether to proceed, ask for more information, re-dispatch with adjusted context, or escalate the task.

Model selection guidance chooses an appropriate model tier based on task scope. It evaluates whether a task is mechanical, requires integration and judgment, or involves architecture and review, then selects the matching model tier for optimal execution.

This workflow is not suited for highly interdependent tasks or undefined plans. It requires a well-defined implementation plan where tasks are mostly independent, as it relies on isolated execution and sequential two-stage review checkpoints to function correctly.