Where Does the Money Go Across Long-Running Coding Agents?

An analysis of a control plane for orchestrating AI coding agents reveals significant and unsustainable costs when scaled to API-based pricing. The primary cost driver was not the implementation or review agents, but a central "controller" agent that remained active for the entire job. This controller's context window grew linearly over time because it was incorrectly persisting and re-reading its own output, such as detailed task briefs for every sub-task. While context compaction was designed into the system, it was being misapplied to short-lived worker agents instead of the long-running controller, leading to massive inefficiencies from re-reading state.