Lore

Two-Tier Model Routing

A production pattern where a small, cheap router model classifies an incoming request and dispatches it to one of two paths: a narrow, highly optimized subprompt for a known, common intent, or the big general-purpose model/prompt for open-ended requests. Linear's agent used this router to send roughly 80% of early use cases — mostly issue creation — to a specialized issue-creation subprompt, reserving the large model for genuinely ambiguous cases. Sizing the model to the task keeps cost and latency down for the majority of requests without sacrificing the flexibility a large model provides for the long tail. Complements Model Tier Selection by Task Upside and Hero-Use-Case Reliability Bar: known hero use cases get narrow, hardened prompts; unclassified variance is routed to the general reasoning path.