Video unavailable. Open the source below.
NewMobile Jev
An Android agent navigates apps with a live studio and execution trace.
Added to Jevfast
How it uses Jev
The mobile agent sends a live device observation, goal, recent actions, and candidate controls to Jev. One Choice selects the next operation; speculative target Choices cover applicable app, tap, scroll, or text targets. The runner consumes only the target branch matching the selected operation, then executes and observes the device again.
What Jev decides
Choose a mobile operation targetExample answers · not a recorded Jev response · Source ↗
Question 1 · tap target
YOUR APP
INSTRUCTION
Assuming the next operation is TAP, choose its best target for the entire goal. This is speculative: another question selects the operation. Use the visible screen and recent actions. Choose only an offered index.
STATE
Tested Settings screen: goal “Open About phone”; foreground package Settings; visible controls Search, About phone, and a scrollable Settings list; text candidate “hello”; no prior action. The target question offers the two observed tappable controls as indices 1 and 2.
JEV · CHOICE
- 1
- 2
App workflow
- Observe the device
The agent captures the screen and installed apps, and carries forward recent actions.
- Build bounded choices
Code proposes available operations and their applicable targets from observed controls, apps, and text candidates.
- Select and validate
Jev chooses an operation and optional target; the policy validates returned choices and consumes only the selected branch.
- Execute and verify
The runner executes against the expected observation, rereads the screen, and handles stale state, input verification, completion, or failure.
Why it is interesting
It makes each UI action inspectable and guards against stale screen decisions, while an execution trace records actions and observations rather than treating a model DONE answer as proof of task completion.