Browser Use + Jev: How a Web Agent Can Complete a Flight Search in Seven Seconds

Browser Agents have two practical problems: they are slow and expensive. They repeatedly read a page, construct an action, call a model, wait, and return the new state. A September 2026 Browser Use + Jev demo reported a flight search in about seven seconds and a disclosed single-run cost of $0.0039.

The original demonstration by Gregor Zunic and the Chinese analysis by AYi are case-study evidence, not a universal performance promise. The public materials did not include a complete repeatable script, network conditions, failure rate, trial count, model version, or cost formula.

The Layered Executor

A conventional loop is:

page screenshot or large DOM
  → general model
  → natural language or tool call
  → browser execution
  → read page again

The Jev design narrows the boundary:

current page state
  → generate valid actions for this step
  → Jev chooses click / select / navigate
  → browser executes
  → small model only when text is needed
  → next state

Visual understanding, state representation, action choice, text generation, and execution do not have to belong to one model.

Dynamic Action Spaces

Exposing every possible browser action makes the model choose among irrelevant links, ads, scripts, popups, scrolling, refresh, and tabs. Instead, construct a state-specific set:

page state S_t
  → identify operable elements
  → filter by task
  → create A_t = {a1, a2, ..., an}
  → choose a_t

An action should carry stable ID, role, accessible name, target reference, allowed arguments, risk level, and page-state version. The runtime can then reject stale or dangerous actions.

DOM First, Not DOM Only

DOM or an Accessibility Tree compactly represents standard forms, labels, buttons, values, disabled state, and loading state. Canvas, maps, charts, overlays, virtual lists, and visual mismatch still require screenshots or vision.

standard form → DOM / Accessibility Tree
visual layout → screenshot / vision model
conflict → pause and recheck state

The goal is to make vision an on-demand capability rather than the only input.

Jev for Actions, Small Model for Text

action selection: Jev → click, choose, navigate, advance state
text entry: small language model → type requested text
execution: browser runtime → validate and perform

This avoids asking a general model to explain every click. It also preserves a clear risk boundary: submission, purchase, message sending, account changes, and uploads require confirmation and audit even if Jev is confident.

Measure the Whole System

Do not repeat “7 seconds” or “450x cheaper” without matching conditions. Measure page load, state extraction, Jev latency, text-model fallback, browser execution, retries, failures, and end-to-end completion.

{
  "task_id": "flight-search-001",
  "success": true,
  "elapsed_ms": 7100,
  "action_failures": 0,
  "text_fallbacks": 1,
  "retry_count": 0,
  "estimated_cost": 0.0039
}

Repeat across pages, accounts, network conditions, and state changes. Report P50/P95 latency, failure recovery, stale-state rejection, and human takeover.

Acceptance Checklist

• Action space is generated from current state and task scope.
• Every action has target, allowed arguments, risk, and state version.
• DOM and visual state conflicts pause the Agent.
• Text fallback is bounded and observable.
• External side effects require confirmation, idempotency, and audit.
• The runtime reads authoritative results instead of trusting the model's claim.
• Public demo numbers are separated from local benchmark results.

Summary

The Browser Use + Jev example is valuable because it changes the architecture, not only the model. A fast typed decision layer handles bounded actions, a small model handles text when needed, and deterministic browser code owns execution and verification.