Video explanation
A model-call meter, phone UI and mechanical actuator visualize an agent learning a test flow, replaying cached actions on subsequent runs, and returning control to the agent when replay no longer works.
Goal-driven actions
The agent selects actions from screen and UI information; a natural-language goal still needs independent checks of the result.
Verified traces can be cached
e2e caches actions that pass recorded checks, allowing replay without repeated model calls for the same step.
Replan when replay fails
A cache is not a permanent guarantee. When UI changes break replay, the agent must find and verify a new path.