The Real Cost of Multi-Agent AI Systems: Context, Not Parallelism

While running four subagents in parallel during a Claude Code session, an engineer stopped to question a nagging assumption: was the cost of delegation simply too many agents? Investigation revealed the biggest expense wasn't the subagents themselves but a routine 'check on the agents' status call that pulled entire raw JSONL transcripts of background agents into the orchestrator's main context — a mistake that repeated twice.

The piece draws a sharp distinction between token spend, which is paid once, and context pollution, which taxes every subsequent turn by competing for the model's attention regardless of window size. It introduces 'cognitive locality': tasks requiring the same mental model should stay grouped, since splitting them forces multiple agents to independently rebuild identical understanding of a codebase.

For engineers building agentic workflows, the takeaway reframes what subagents are actually for — not raw parallel speed, but isolating noisy intermediate reasoning so it never pollutes the orchestrator's working memory. The author is careful to flag this as a working hypothesis from self-reported orchestrator analysis, not measured instrumentation.