Use explicit, bounded subagents to answer independent codebase questions and summarize their evidence.
