Google's ARTEMIS Hits 99% on Android Tasks Where Most Agents Fail

Google's Pixel Test Engineering Fusion team just released ARTEMIS, an open-source Android automation agent that turns plain English instructions into reliable device control. It plugs into AI coding assistants like Antigravity, Codex, and Claude Code through the Model Context Protocol, letting you drive a real phone from your IDE chat window. The project claims a 99%+ success rate on AndroidWorld, Google Research's 100+ task benchmark that most mobile agents struggle to clear.

Why phone agents keep falling over

Two problems have kept language-model-driven phone automation from being useful in practice. The first is latency. Conventional mobile agents rely on a fully synchronous blocking model where every action waits for the LLM to prune history, check assertions, and audit long-term plans, inflating per-step latency to 20-40 seconds. The second is fragility. System popups, permission dialogs, and transient UI controls (think video player overlays that fade after two seconds) constantly break scripted flows.

ARTEMIS attacks both. Borrowing from optimistic concurrency control in database systems, it decouples the main execution loop from heavy auxiliary computation. The front-facing path narrows to a high-speed perception, decision, safety gate, execution pipeline, while context pruning, milestone checkers, and planner validations run in parallel without halting device interaction. When background verification detects a deviation, ARTEMIS rolls back via state snapshots and injects self-healing feedback. Per-step latency lands in the 3-5 second range in Flash mode.