When Claude Code skills or prompts orchestrate multiple background subagents, a subtle but devastating problem emerges: context overflow from agent transcripts. This guide explains the problem, when it matters, and the battle-tested solution.
When you launch a background subagent with the Task tool (run_in_background: true), the natural instinct is to call TaskOutput to collect its results. But TaskOutput doesn't return just the agent's final answer — it returns the entire agent transcript: every web page fetched, every search result, every tool call, every reasoning step.