Subagent delegation

Subagents run with their own context window. They are the correct tool when the work is searchable, parallelizable, or produces output you only need summarized.

Delegate these

• Wide codebase searches where you need locations, not full file contents.
• Independent implementation tasks with a crisp spec and a test gate.
• Reviews against a checklist, returning findings rather than edits.
• Any task whose raw output would flood your session with tokens you will not reuse.

Do not delegate

• Decisions that need the full conversation history.
• Work with unclear acceptance criteria; a vague brief returns confident wrong code.
• Anything touching credentials or secrets handling.

Write briefs that survive handoff

A good brief states the goal, the constraints, the files in play, and how to verify success. Include the exact command that proves it works. If the agent cannot verify its own output, you have written a wish, not a brief.

Trust but pin

Ask for a diff or a summary of changed files, then review those files yourself. Subagent confidence is not correctness. The verification step is yours.