The Orchestrator's Tax

martinfowler.com · 2 months ago · 48

A developer working with Claude Code on multi-agent tasks discovered that the main cost wasn't running four parallel subagents, but rather the orchestrator agent polling for status updates and dumping full transcript data back into its context, which polluted its working memory for subsequent decisions. The largest avoidable expense came from two transcript-dump operations that brought tens of thousands of tokens of raw JSONL and intermediate reasoning into the main thread unnecessarily. The developer concluded that subagents are most valuable not for raw parallelism, but for isolating noisy reasoning from the orchestrator's context, and encoded this insight into standing rules about cognitive locality—keeping tasks with shared mental models together, avoiding unnecessary polling, and preventing repository-wide operations across concurrent agents.