Subagent-Driven Development
Intent
Execute plan tasks via focused subagents with explicit review loops to reduce context pollution and improve quality.
When to Use
- •Tasks are independent and can be executed in isolation.
- •You want faster iteration with structured review between tasks.
Precondition Failure Signal
- •Multiple subagents edit overlapping files simultaneously.
- •Tasks proceed without review or verification.
- •Subagents lack full task context.
Postcondition Success Signal
- •Each task is executed by a fresh subagent with clear scope.
- •Spec compliance review occurs before code-quality review.
- •Plan status and evidence are updated after each task.
Process
- •Prepare: Extract full task text and context from the plan.
- •Dispatch: Send a single task to one subagent with constraints and expected output.
- •Review (Spec): Validate that the task meets plan requirements.
- •Review (Quality): Validate code quality and standards.
- •Update Plan: Record verification evidence and mark the task Done.
- •Repeat: Continue task by task until complete.
- •Finalize: Use
finishing-a-development-branch.
Example Test / Validation
- •Task shows evidence from tests and both review stages before being marked Done.
Common Red Flags / Guardrail Violations
- •Skipping spec review.
- •Allowing subagents to read the plan instead of providing full task context.
- •Running multiple implementation subagents in parallel for the same area.
Recommended Review Personas
- •Tech Lead - validates spec compliance.
- •Software Engineer - validates code quality and consistency.
Skill Priority
P3 - Delivery & Flow
Conflict Resolution Rules
- •If a review fails, the same task must be fixed before moving on.
- •Quality and verification skills override speed.
Conceptual Dependencies
- •requesting-code-review
- •verification-and-handover
Classification
Delivery
Operational
Notes
Fresh context per task improves quality and reduces drift.