A developer reflects on a real Claude Code session involving four subagents on a .NET codebase refactor, discovering that the biggest cost wasn't parallelism overhead but orchestrator context pollution. When status-check tool calls pulled full raw transcripts (tens of thousands of tokens) into the main thread, those tokens persisted across every subsequent turn, degrading the orchestrator's working memory. The author introduces the concept of 'cognitive locality' — tasks sharing the same mental model should stay together rather than be split across agents — and argues that subagents' primary value is keeping noisy intermediate reasoning out of the orchestrator's context, not just saving wall-clock time. The post concludes with four concrete CLAUDE.md rules (limit wave size, avoid transcript polling, ban concurrent repo-wide git ops, treat overlapping file ownership as a consolidation signal) and a heuristic for when to add standing instructions versus when a single clarifying fact suffices.
