“Every token in the orchestrator’s context is competing for its attention, and the real value of a subagent is what it keeps out of that context, not how fast it runs.”
Multi-agent setups get sold on parallelism. The actual constraint is that the coordinating agent’s context fills with junk from its subagents and every decision after that gets worse. Tokens are a one-time cost. A polluted context is a cost you pay on every turn that follows. Rare to see someone write about agent architecture without trying to sell a framework at the end.