Description
Writing end-to-end Android tests has always been tiring and brittle: element coordinates, wait timing and every if-branch by hand, and a screen redesign breaks the script wholesale so you go back and fix each one. Google's open-source ARTEMIS takes another approach — you describe the flow in natural language and it turns that into reliable Android automation, planning the steps itself, running them on a real device or emulator, and capturing the logs and results.
It is not a record-and-replay brittle script. Given "open Settings, turn Wi-Fi off then on," it produces a plan and executes it step by step, keeping logs, screenshots and a report along the way, so a small UI change does not collapse the whole case. It reaches a 99%+ success rate on the public AndroidWorld benchmark. More usefully, it plugs straight into AI coding assistants like Antigravity, Codex and Claude Code — so beyond writing code, an agent can also take over "run it on a real device and see if it works."
Written in Python, Apache-2.0.
Natural language to automation: describe a flow in a sentence and it becomes reliable Android actions, with the steps planned and run on a real device or emulator.
Plan then execute: an instruction yields a plan that runs step by step, rather than a brittle record-and-replay script, so a small UI change does not break the whole case.
Logs and reports: the run captures logs, screenshots and a result report, so a failure shows exactly which step stalled.
AI-assistant integration: it slots into Antigravity, Codex and Claude Code, letting an agent take over running the check on a real device.
High success rate: 99%+ task success on the public AndroidWorld benchmark.
It is not a record-and-replay brittle script. Given "open Settings, turn Wi-Fi off then on," it produces a plan and executes it step by step, keeping logs, screenshots and a report along the way, so a small UI change does not collapse the whole case. It reaches a 99%+ success rate on the public AndroidWorld benchmark. More usefully, it plugs straight into AI coding assistants like Antigravity, Codex and Claude Code — so beyond writing code, an agent can also take over "run it on a real device and see if it works."
Written in Python, Apache-2.0.
Features
Natural language to automation: describe a flow in a sentence and it becomes reliable Android actions, with the steps planned and run on a real device or emulator.
Plan then execute: an instruction yields a plan that runs step by step, rather than a brittle record-and-replay script, so a small UI change does not break the whole case.
Logs and reports: the run captures logs, screenshots and a result report, so a failure shows exactly which step stalled.
AI-assistant integration: it slots into Antigravity, Codex and Claude Code, letting an agent take over running the check on a real device.
High success rate: 99%+ task success on the public AndroidWorld benchmark.
