LLM routing aims to pick a capable model at low cost, yet specialized routers often need updates as models and tasks evolve. To address this, we introduce COMPASS, an agentic routing framework. An LLM follows a routing recipe, gathers performance and cost evidence via tools, and can inspect candidate answers before deciding which model to invoke.\ \ Evaluations on LLMRouterBench and DeepSWE demonstrate that agentic routing can improve accuracy and reduce model‑call cost without any routing‑specific training. Specifically, COMPASS reaches 69.85% accuracy on LLMRouterBench at a cost of $139.57, compared with RLCascadeRouter’s 69.14% at $180.22. On DeepSWE under fixed five‑fold cross‑validation, it attains 78.1% accuracy at $215, whereas the best single model achieves 74.1% at $500.\ \ Moreover, COMPASS supports routing across benchmarks and adapts to an expanded model pool simply by reading new evaluation records, without changing the router’s weights or prompt.\ \ Review: COMPASS showcases the promise of capability‑oriented profiling combined with tool interaction for adaptive routing, delivering efficient model selection while sidestepping the maintenance overhead of frequent router retraining.