Agent Device, a CLI tool for AI-driven mobile app automation, achieved over 50% reduction in LLM token consumption and a 33% speed improvement through several optimizations. Key changes include using accessibility tree snapshots instead of screenshots (3.5x efficiency gain), aggressive trimming of off-screen UI elements (up to 80% token reduction in long lists), improved agent skill instructions to reduce unnecessary steps, and Android-specific reliability fixes. A complex 9-minute, 150k-token debugging session now completes in 6 minutes using ~69k tokens. Future plans include diff snapshots, further skill schema refinement, and log analysis efficiency improvements.
Table of contents
Why we talk about token efficiencyWhat is Agent DeviceHow we made our AI Agent cut token usage in halfWhat’s next?Try it outSort: