Jev Makes Browser Agents Choose, Not Write

Jev’s strongest idea is architectural: reserve generative models for language and use fast, typed decisions for repetitive interaction loops. The demo is compelling, but its speed claim remains a narrow showcase rather than a general browser-agent benchmark.

• –Indexed DOM actions reduce selector hallucinations and keep model output away from executable code.
• –One Jev request can choose both the operation and compatible target, reducing network overhead.
• –A separate text model preserves flexibility for form entry without burdening Jev with prose generation.
• –The approach depends on structured, accessible browser state and is less suitable for canvas-heavy or visually complex interfaces.
• –Jev is hosted and proprietary, so the open-source agent does not remove inference-provider dependence.