Agent-as-a-Router: Agentic Model Routing for Coding Tasks

The article introduces "Agent-as-a-Router," a novel framework for routing coding tasks to the most suitable Large Language Models (LLMs) by addressing the information deficit in existing static routers. The framework, instantiated as ACRouter, employs a C-A-F loop (Context->Action->Feedback->Context) and includes components like an Orchestrator and a Verifier, achieving a 15.3% performance improvement over heuristic routers. It is evaluated using CodeRouterBench, featuring approximately 10,000 task instances from eight leading LLMs, demonstrating superior performance in minimizing cumulative regret on both in-distribution and out-of-distribution tasks, which is critical for optimizing resource allocation in multi-LLM environments.