Google Releases Open-Source ARTEMIS for Controlling Android Phones with AI Agents... Achieves 99% Completion Rate on AndroidWorld

Google has released ARTEMIS, an open-source project that enables AI agents and test suites to operate real Android devices using natural language-based commands. This tool performs cross-app automation and recorded a completion rate of over 99% on Google Research's AndroidWorld benchmark (over 100 multi-step tasks).

Key Features and Architecture

ARTEMIS provides two main profiles:

• Flash Profile: A reactive loop where a single model observes the screen and acts without graph orchestration, offering fast response times of approximately 3–5 seconds. It manages context by compressing past screenshots into visual summaries.
• Pro Profile: A multi-agent graph-based workflow that takes approximately 15–40 seconds, where a Planner establishes the plan and an Operator executes it. It performs Pre-Execution Checks before each action, and the Operator handles recovery in case of failure. It supports long-term workflows of over 100 steps.

UI element localization combines the accessibility layer, OCR, and vision models to support various UI frameworks such as Canvas, Compose, and Flutter.

Developer Integration

ARTEMIS includes a native Model Context Protocol (MCP) server, allowing integration with major AI IDEs and coding assistants such as Antigravity, Claude Code, and Windsurf. This enables bug reproduction and test execution on real devices, and it can also be integrated into pytest or CI/CD pipelines via the Python SDK.