What's this about?
GUI agents that drive Android phones are usually tested in clean environments. AnTrap builds a dynamic adversarial environment instead — unexpected pop-ups, permission dialogs, network failures, layout changes mid-task — and evaluates whether agents stay on task. They largely don't.
What is a "GUI agent"?
An agent that operates software the way a person does: it looks at the screen, decides where to tap, taps, and looks again. It has no API, no structured state, only pixels and its own memory of what it was doing. That makes it extremely general — it works with any app, including ones with no automation interface — and extremely fragile, because a modal dialog that a human dismisses without thinking can send it down a completely wrong path.
Why does this matter to you?
- "Robust against runtime anomalies" is the property that separates a demo from something you'd let touch a real account.
- The adversarial-environment framing is the right one: the interesting failures come from the environment changing under the agent, not from bad prompts.
- Directly relevant if any part of your scanning or automation runs against a UI rather than an API.