Google Open-Sources Android Automation Agent Project ARTEMIS, Achieving Over 99% Completion Rate on AndroidWorld
Google has open-sourced the Android automation agent project ARTEMIS. Users issue tasks in natural language, and the agent on a real device views the screen, finds controls, swipes and taps, and brings back screenshots and logs. The project uses the Apache-2.0 license and is implemented in Python, comes with a native MCP Server, can connect to IDEs such as Claude Code, Codex, and Windsurf, and supports models including Gemini, Claude, GPT-4o, and Qwen-VL. Google says it achieves a task completion rate of over 99% on Google Research's AndroidWorld benchmark, and distinguishes between two execution modes: the fast Flash mode and the slower Pro mode.
Why read this
SelectedAI score 78
The article breaks down how ARTEMIS locates controls through three approaches—accessibility hierarchy, coordinates, and vision—as well as the trade-offs between the Flash and Pro execution modes.
AI Trends