Model Routing

Pain Signals — You Need This Pattern When:

• Most turns are simple but a few need top-tier reasoning; using the top model for everything is expensive
• Some turns are latency-critical (user typing) and others are not (background analysis)
• Cost per session varies wildly and most of it goes to easy turns that didn't need premium tiers
• Different agent roles have different needs (planner = capability, executor = cost) — same model is wrong for both
• You have multiple providers / models and want to use whichever is cheapest at adequate quality

Routing is a proactive choice; graceful-degradation is reactive. Routing picks the right model up-front; degradation switches when the chosen model fails. Both layers needed in serious harnesses.

Core Principle

A router is a function (turn_context) → model. It runs before the model call. The choice depends on signals available at that point — and those signals must be cheap to compute.