Video unavailable. Open the source below.
NewJev Ultrafast
A browser agent selects actions from the current DOM and demonstrates flight search.
Added to Jevfast
How it uses Jev
At each decision step, the agent sends Jev the current page text, numbered controls, recent actions, and the user's goal. One request asks for an operation Choice and operation-specific target Choice questions; code uses only the target matching the selected operation. If the operation is TYPE_TEXT, a separate small LLM writes the field value.
What Jev decides
Jev request and response2 of 3 questions shown
Question 1 · operation
YOUR APP
INSTRUCTION
Advance the one-way Zurich to London flight search from the current page. Choose one operation.
STATE
Google Flights home page. Ticket type: Round trip; one adult; Economy. Origin, destination, departure, and return fields are empty. No recent actions.
JEV · CHOICE
- CLICK 0.77
- TYPE_TEXT 0.22
- BLOCKED 0.01
- SCROLL_DOWN 0.00
- WAIT 0.00
- DONE 0.00
Question 2 · click target
YOUR APP
INSTRUCTION
If the operation is CLICK, choose the observed control that best advances the flight search.
STATE
Google Flights home page. Ticket type: Round trip; one adult; Economy. Origin, destination, departure, and return fields are empty. No recent actions.
JEV · CHOICE
- [12] Change ticket type. Round trip 0.88
- [15] Open Where from? 0.12
2 of 25 offered options shown
App workflow
- Observe page
Read visible text and numbered controls.
- Ask Jev
Send the shared state with operation and target questions.
- Use the answer
Read CLICK and its selected [12] Round trip target.
- Check and act
Validate the observed target and current page, then click.
- Observe again
Read the changed page before the next decision.
Why it is interesting
Jev chooses among actions the page actually exposes, while code keeps browser execution and text entry bounded. Judging the operation and possible targets in one request gives the agent a reusable pattern for changing interfaces without a second Jev round trip.