Developer routes coding agent phases to different AI models to cut cost and noise

A developer has built a tiered routing method called pi-tiered-router that assigns different AI models to distinct phases of a coding task — planning, validation, execution, and tool parsing. The approach avoids using expensive frontier models for low-difficulty work like compressing tool output or classifying task complexity, which are instead handled by cheaper models like Claude Haiku. Planning and validation happen outside the executor's context, so the coding model only sees a clean, distilled plan rather than the full reasoning trace. The system also tiers tasks by complexity, escalating model effort for harder requests while bypassing planning and validation entirely for trivial ones. The developer emphasizes this is not a cost-cutting tool but a quality-matching system, ensuring each phase always gets the model best suited to produce accurate results.