Most LLM routing today is just traffic control: pick a model, retry on failure, track costs. That approach is becoming obsolete. A new concept called second-half routing treats the router as a reasoning agent, orchestrating decisions without changing model weights.

This moves beyond simple proxy logic into four key patterns. First, semantic routing uses a cheap classifier to analyze task intent, domain, and difficulty before choosing a strategy. Second, strategy trees enable search and backtracking, exploring multiple paths and evaluating partial results—essentially applying Tree of Thoughts to routing.

Third, reflexive routing creates a feedback loop where the system captures performance data, generates reflections, and updates its own policy for online learning. Finally, cross-LLM collaboration coordinates a team of specialized experts—classifier, reasoner, retriever, and verifier—to produce a final, verified answer. The router becomes a true intelligence control plane, not just a dumb dispatcher.

This is critical for cost-optimized, safety-conscious, and multi-tool agent systems.