What is actually limiting model routers for coding tasks? Most routers treat picking a model as a static, one-off classification. This paper identifies the real bottleneck as information deficit. Simp

What is actually limiting model routers for coding tasks? Most routers treat picking a model as a static, one-off classification. This paper identifies the real bottleneck as information deficit. Simply augmenting a vanilla LLM router with task-dimension-level performance statistics yields a 15.3% relative gain, surpassing a heuristic router built on the same priors. Motivated by that, the authors propose Agent-as-a-Router, which formalizes routing as a Context to Action to Feedback to Context loop. It closes the information gap by accumulating execution-grounded experience during deployment, so the router improves from real outcomes instead of guessing once. Why does it matter? If you orchestrate multiple coding models by cost and skill, treating routing as a learning loop rather than a one-shot classifier is the lever. This means that routing becomes an agentic process that updates its priors as it runs. Paper: Learn to build effective AI agents in our academy