One month ago, I created Laya, now I raced two agents against multiple use cases. The first one had an LLM inside the orchestrator, and the other had a decision model. The decision model one is 2 times faster than the LLM one and reduces token usage by 80%, even with improved quality judged by Opus blindly.