Engineer's AI-assisted refactor reveals hidden costs in agent orchestration

Software developer Aashish Bhandari used two AI coding agents, Goku and Naruto, to refactor ReviewWithAI, a Markdown document review application, across a two-day engineering session in September 2026. The project spanned eleven low-level designs covering server structure, authorization, testing, and release preparation, with Bhandari overseeing key decisions and approvals. Post-implementation analysis of telemetry data found that repeated short waits by the parent agent for delegated workers consumed over a quarter of its recorded tokens. The study also flagged that worker conversation continuity may improve output quality but raises context costs, a tradeoff that requires controlled testing to resolve. Researchers noted significant gaps in their own measurement tools, concluding that agent orchestration mechanisms need rigorous evaluation alongside code output before meaningful optimization can be claimed.