# Orchestrator tax According to Rahul Garg (Thoughtworks, on martinfowler.com), the real value of a subagent is what it keeps OUT of the orchestrator's context, not how fast it runs. Context isolation is more valuable than parallelism. The "tax" is the pollution that accumulates in the orchestrator's context (i.e., working memory) over a long session. Tokens are spent once, but polluted context degrades EVERY subsequent turn. Bigger [[Context Window|Context windows]] don't fix this; they just let more noise pile up. He also names a useful principle, **cognitive locality**: tasks that need the same mental model belong in the same agent. Split too finely, and each agent needs to rebuild the same understanding from scratch, with the risk of deviating from session to session. His working rules: - Two to four agents per wave; if you think you need five or more, consolidate. - Don't poll for status, and never pull full transcripts for lightweight questions. - Keep repository-wide git operations out of concurrent prompts (overlapping file ownership hurts). - Encode clarifying facts in standing rules (CLAUDE.md), not process procedures. This matches my experience with [[Claude Code]] subagents and connects directly to [[Context Engineering]]; what matters is keeping a clean context. ## References - The Orchestrator's Tax (Rahul Garg): https://martinfowler.com/articles/orchestrator-tax.html ## Related - [[Context Engineering]] - [[Context Window]] - [[AI Agents]] - [[Claude Code]]