orchestrator

To maintain context-window hygiene during multi-step coding workflows, you can decompose requests and delegate sub-tasks to subagents. This prevents large tool outputs and file dumps from polluting the main conversation, returning only clean summaries.

Task decomposition for code reviews involves breaking a large request into discrete sub-tasks and sending them to focused subagents. The subagents execute the review scoped to their prompt and synthesize their outputs into a single clean report.

You should delegate work to subagents for multi-step workflows requiring sequential or parallel tool operations, such as debugging, refactoring, or research. Acting directly is better for simple decisions where heavy tool chatter and file dumps are not expected.

Yes, you can use subagents for parallel file editing and research tasks. They execute the scoped work independently and return only the necessary synthesized results, ensuring you receive the final outputs without the raw tool noise.

The best way to automate complex refactoring workflows without clutter is to use an orchestration approach that delegates the work to subagents. This forces task decomposition and requires subagents to summarize results, keeping the main context focused on outcomes.