DynBranch is a speculative subgraph reuse technique for dynamic agentic LLM serving. It reduces latency by up to 32% and 46-66% by reusing completed subgraph results across requests, without requiring changes to agent harnesses or model execution engines.