Every AutomateAI agent works the same way: look at the screen, decide what a person would do next, act, and check the result. The only thing that changes is what kind of screen it's looking at.
The web agent reads the actual rendered page — not a fixed selector map — and decides what to click, fill, or wait for based on the goal it's given.
Runs against your own device farm, tracking app state through auth, deep links, and the kind of flaky navigation that only shows up on a real handset.
Instead of sending isolated messages, the chatbot agent carries context across turns the way an actual user would — then checks whether the conversation stayed correct and on-track.
Since the agent is reading the rendered screen for every other test anyway, it runs a structural pass, a vision pass, and a cross-check between the two — mapped to the standards your compliance team tracks against.
The agent already knows when it tapped something and watches the screen directly, so it can flag real response-time regressions without an APM SDK anywhere in your codebase.
Agents work within a fixed step budget per run, so a confused agent fails cleanly with a reason instead of looping forever or running up cost.