What we're learning building AppClaw, and where AI is taking mobile testing, from how the agent reads a screen to when a plain-English test beats a brittle selector.
The honest worry: if an LLM drives every step, every run costs money. It doesn't have to. Here's what actually happens behind each step, why the model in DOM mode only parses language and never sees the screen, and how AppClaw caches and heals locators so repeat runs skip it entirely.
AppClaw reads the structured accessibility tree when it exists (fast, exact, cheap) and falls back to vision on the raw pixels when it doesn't. Here's how the agent decides, and why the same tap works either way.
Selectors describe how an app is built. Instructions describe what you want. AppClaw closes the gap by resolving intent on the live screen instead of matching a fixed path.