subagent-driven-development

Context pollution occurs when task history accumulates in a single agent's memory, degrading focus and quality. Subagent-driven development dispatches a fresh AI agent for each task, isolating context and ensuring each task receives undivided attention without prior task artifacts interfering.

Enforce mandatory code review after every task completes. This catches issues early before they compound into later tasks, maintains quality throughout implementation, and enables rapid iteration by catching defects at task boundaries rather than at final review.

Use subagent-driven development to dispatch a dedicated subagent per task, request code review after each task, apply fixes, then move to the next task. This workflow keeps all execution within one session while maintaining quality gates and preventing context degradation.

Subagent-driven development optimizes plan execution when tasks are largely independent. For tightly coupled or sequential tasks with heavy dependencies, a single continuous agent may be more efficient, as fresh subagent dispatch adds overhead without context reuse benefit.

You need a loaded implementation plan that breaks work into discrete, largely independent tasks. The plan should define task scope, success criteria, and any shared context or outputs tasks reference, allowing subagents to execute with minimal cross-task coordination.

Fresh subagents eliminate accumulated context drift and fatigue that degrades decision-making in long sessions. Paired with mandatory code review between tasks, this separation ensures each task receives focused reasoning and early defect detection before issues cascade downstream.