subagent-driven-development

To execute implementation plans, you dispatch fresh subagents per task using prompt templates. Each subagent operates with isolated context to prevent leakage, running through implementer, spec reviewer, and code quality reviewer stages.

A two-stage code review workflow checks spec compliance first, then evaluates code quality. This ensures the implementation matches requirements before verifying maintainability and correctness through dedicated reviewer subagents.

Automated plan execution handles blocked tasks through structured escalation statuses like BLOCKED or NEEDS_CONTEXT. The workflow pauses specific tasks requiring human input while allowing other isolated subagents to continue progressing.

Yes, you can run implementation tasks in parallel within a single session. The workflow supports optional parallel execution by dispatching multiple fresh subagents simultaneously while maintaining strict progression rules.

Context leakage happens when shared state contaminates task execution. Subagents prevent this by operating with isolated context, ensuring each task gets its own implementer and reviewers without prior session interference.

Continuous execution workflows without human-in-the-loop review rely entirely on automated spec and code quality checks. Limitations include potential misinterpretation of ambiguous requirements, necessitating BLOCKED or NEEDS_CONTEXT escalations for complex scenarios.