A new architectural approach introduces an agent runtime layer for multi-agent LLM systems, enhancing the interaction between agent frameworks and serving engines. This change has shown promising results, including improved cache hit rates and reduced latency in real workloads.