The Orchestration Tax

This article discusses the concept of 'orchestration tax' in AI agent development, where running multiple agents does not scale human productivity due to cognitive bandwidth limits. It compares the human developer to a serial processor (like Python's GIL) in a concurrent system, using Amdahl's Law to explain why parallel agent execution hits a hard cap. The author argues that managing many agents requires architecting one's attention, not just adding more agents, and emphasizes that feeling busy is not the same as being productive. The piece draws from a panel at Google I/O and offers insights for software engineers navigating agentic workflows.