WIP: AMAS: An Android Macro Action System for Efficient LLM Mobile Agents

Abstract

Recent advances in LLMs have enabled mobile agents that can automate user tasks through natural language interaction with app interfaces. However, current agents execute tasks as sequences of basic UI actions (taps, swipes), incurring extra latency, token cost, and accumulating error risks. Motivated by this, we propose AMAS, an Android Macro Action System that constructs reusable, system-grounded macro actions by identifying system-level application entry transitions that directly jump to task-relevant, app-specific screens. Results show that AMAS reduces agent's average LLM calls by 23.22%, LLM latency by 26.38%, and token cost by 21.9%, while having a 6.0% gain on success rate. Further coverage analysis shows that AMAS constructs 184 stable and distinct macro actions across the 20 benchmark apps, averaging 9.2 per app. At the task level, AMAS provides every task with at least one macro action, and 54.3% of tasks coverage. These results demonstrate that system-level macro actions can provide broad coverage while improving both the efficiency and reliability of LLM-based mobile agents.