Mobile GUI agents often need to carry out repetitive tasks with similar interaction flows. Reusing previous executions can cut down LLM inference calls and runtime overhead, yet variations in user instructions or execution environments may invalidate earlier successful actions. We categorize such variations into six representative types and observe that not every step must be redone: individual transitions remain reusable as long as they are validated against the current task context and UI state.\ From this insight we introduce graph‑text templates for selective trace reuse. The graph captures UI states and action transitions, while the accompanying text encodes task intent, procedural structure, and typed slots for variable content.\ Our record‑template‑replay framework first builds templates from verified executions, then leverages the incoming instruction and live UI state to match, adapt, or discard each transition.\ Evaluation on the AndroidWorld benchmark shows that token consumption for LLM inference drops by roughly 80% on average across five base agents, and most agents achieve at least a 1.5× speedup compared to LLM‑only execution. Real‑device tests on Android and HarmonyOS report action reuse ratios of 70.7%‑79.8% and more than a 2× speedup in most configurations.\
Review