Jev Ultrafast: A browser agent with a dynamic, indexed action space

Jev Ultrafast is a browser agent that drives web interaction by converting each DOM snapshot into an indexed element table and offering a compact, compatible action space (operations: CLICK, TYPE_TEXT, SELECT, SCROLL_UP, SCROLL_DOWN, WAIT, DONE, BLOCKED). Every decision cycle is one network round trip: the agent picks an operation and a target head from the observed state, and a small LLM is invoked only when TYPE_TEXT to produce typed content. Targets are restricted to compatible elements and native dropdown choices use observed option indices. Executed targets are resolved from observed DOM nodes, revalidated for freshness and occlusion, and model output is never executed as selectors, coordinates, or JavaScript; text-helper output must parse as a small JSON object before typing. The loop waits briefly for useful post-input state (e.g., up to 200 ms for combobox suggestions, two animation frames or 50 ms for other interactions) and keeps hidden tabs rendering via focus emulation.

Measured runs show a real Google Flights task completing in 7.073 seconds including model calls and loading; in six alternating runs median task time dropped from 9.450 s to 7.092 s (25% reduction) and median browser protocol calls fell from 1,092 to 101. The same policy opened a Wikipedia page in 2.798 s and passed a local hotel search in 1.896 s. The codebase provides an inspector, demo scripts, and modular components (agent loop, atomic DOM snapshot, browser executor, model interface). Current limitations include partial accessibility support, missing handling for shadow roots/frames/canvas/uploads/pop-ups and complex widgets; DONE still requires independent outcome verification.