Why AI coding orchestrators hit hard limits no better code can fix

A software development workflow can be managed by an 'orchestrator' skill that directs AI coding agents through ordered phases, but this approach has fundamental constraints tied to where the procedure runs, not how well it is written. The orchestrator operates inside the developer's own agent session, meaning it inherits whatever model, reasoning tier, and AI provider that session uses, with no ability to self-specify cost or capability. Because the orchestrator runs within a single conversation, it cannot prune its own context window, execute two phases simultaneously, or hand off steps that require no AI reasoning. Phase order is enforced only by the model's inclination to follow prose instructions, not by any external programmatic control. A separate 'harness' approach — moving the procedure into an external program that calls the coding agent rather than a document the agent reads — is proposed as a potential solution, with its trade-offs to be examined across five subsequent parts.