This paper explores the tradeoffs between latency, reliability, and cost in workflows that integrate large language models (LLMs) with traditional computational agents. It presents models and strategies for optimizing these workflows, which could enhance the efficiency of AI systems in various applications.