Summary: Jev Ultrafast

Jev Ultrafast is a specialized browser agent designed to optimize the efficiency and speed of automated web interactions through a dynamic, indexed action space. By abstracting the browser interface into a structured table of elements, it moves away from traditional screenshot-based navigation, resulting in significantly reduced latency and computational overhead.

#### Core Methodology
The agent operates on a continuous feedback loop consisting of state observation, action prediction, and execution. Instead of analyzing pixel-based screenshots, Jev generates an atomic snapshot of the DOM, indexing accessible controls (e.g., buttons, comboboxes, textboxes) with numerical identifiers.

1. Action Space and Decision Making:

The agent employs a "TypeSafe" request architecture where each state observation produces a table of elements paired with compatible operations. The decision-making process is bifurcated into two parallel "heads":

• Action/Target Head: Selects the operation (e.g., CLICK, TYPE_TEXT, SELECT, SCROLL) and the specific element index from the current DOM table.
• Text Generation Head: A small LLM is invoked strictly when the TYPE_TEXT operation is selected to generate the required input content.

This design limits the model's output to valid, predefined operations and targets, preventing the generation of arbitrary selectors or executable code.

2. Execution and Efficiency:

Jev minimizes browser overhead by maintaining state atomicity—one request per decision cycle. It performs "speculative" target resolution where it validates control geometry and visibility before execution. If a control is covered or hidden, it is excluded from the prediction phase.

• Dynamic Wait Logic: The agent manages timing by waiting for useful state transitions (e.g., waiting for suggestion lists after typing in a combobox) capped at 200ms, while other interactions are constrained to 50ms or two animation frames.
• DOM Freshness Guards: The executor checks page freshness and occlusion for every target, ensuring the browser state remains consistent with the model’s observations.

3. Technical Performance:

By removing the reliance on vision-based models for spatial reasoning, Jev achieves high-speed navigation. Performance benchmarks indicate a 25% reduction in median task time (9.450s to 7.092s) and a substantial decrease in browser protocol calls (from $\approx 1,092$ to $101$) compared to baseline approaches.

#### Constraints and Scope
While highly efficient, the current iteration is an MVP (Minimum Viable Product). It handles standard HTML and ARIA controls but does not support:

• Shadow DOMs or nested frames.
• Canvas-based elements or arbitrary custom keyboard widgets.
• Complex upload workflows or multi-tab management beyond current profile constraints.

The system relies on an OpenAI-compatible API for its text-helper model, allowing for flexible backends (e.g., Gemini, DeepSeek, or Mercury-2.5) while emphasizing that the model itself never interprets the page visually; it only interacts with the textual, indexed representation of the DOM.