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

Who Should Read

A developer whose browser agent is slow and expensive and who wants to run it on structured DOM state instead of screenshots

Core Mechanics

• Page elements become a numbered table
  • Every observation builds a fresh table like [1] button, [2] combobox.
  • The operations are CLICK, TYPE_TEXT, SELECT, SCROLL_UP, SCROLL_DOWN, WAIT, DONE, and BLOCKED.
  • Only supported operations and targets are shown to the model.
• Operation and target in one request
  • The operation head and target heads read the same state in one network round trip.
  • Target questions are speculative, so if the operation is CLICK only click_target executes.
  • A small LLM writes text only for TYPE_TEXT.
• Structured state, no screenshots
  • The default loop uses no screenshots.
  • Only visible text is sent, so offscreen bodies and footers do not fill the context.
  • One browser call per snapshot reads control names, values, and text atomically.
• Model output never becomes code
  • Every executed target is resolved from an observed node.

Community Reactions

• Opinion that a cloud model is disappointing
  • Commenters wanted a model that runs fast on consumer hardware.
  • One did not want to pay an AI tax to gatekeepers on every use.
• Experience raising telemetry concerns
  • One commenter inspected browser-use and said default-on PostHog telemetry is excessive.
  • They claimed browser-harness can leak credentials.
• Question about the timing boundary
  • A commenter asked whether the initial page observation, excluded from timing, is the slowest part.
• Reports that it does not run
  • One person got an error on the run demo button, with no build error.
  • Another simply said it does not work.
• Opinion that Google Flights needs no agent
  • A commenter said the flight URL can be constructed with protobuf.

How to Apply

• Try the numbered table with a prompt
  • Convert your agent's DOM dump into lines like [id] role name · value.
  • Restrict the model's reply to an operation and an id in JSON.
  • The prompt below lets you test it in 30 minutes.
• Remove screenshots from the default loop
  • Send only visible control text and values as a structured snapshot.
  • Turn screenshots on only for debugging.
• Limit model output to node ids
  • Accept only observed element ids instead of selectors or coordinates.
  • Re-resolve the node and check for page changes and covered elements right before acting.
• Add independent verification after DONE
  • Even when the agent says it is finished, recheck route, date, and results in code.

Try It

# Paste into Claude or ChatGPT (simplified version of the Jev approach: pick an action from a numbered element table)

You are the policy model of a browser agent. Pick exactly one next action from the element table below.
Allowed operations: CLICK, TYPE_TEXT, SELECT, SCROLL_UP, SCROLL_DOWN, WAIT, DONE, BLOCKED
Reply with one line of JSON only: {"operation":"...","target":number or null}
Do not use ids that are not in the table, and no selectors or coordinates.

Goal: {{goal}}

Element table:
{{element table, e.g. [1] button Change ticket type · Round trip
[2] combobox Where from? · San Francisco
[3] combobox Where to? · empty
[4] textbox Departure · empty}}

Next to try: when TYPE_TEXT comes back, generate the text in a separate request with a small model and compare response speed.

Terminology

Related Papers

• Same Simulation, Different Evidence: Where Retrieved Tool Output Costs LLM Agents Accuracy

  Saving long tool output to files cut GPT-5.1 accuracy by 11.3 points; naming the authoritative file restored it.

• Fake News, Real Crashes: Information Injection Attacks on LLM Financial Trading Agents

  Fake headlines hijacked LLM trading agents 62% of the time versus 14% for prompt injection.

• Harness Engineering 101: The LLM is the brain, the harness is the body

  An agent is a loop managing a JSON array; 16 chapters build one from scratch

• Harness2 by Google Research – Recursive Agent Harnessing

  Recursively refining the harness with fixed model weights raised benchmark scores by up to 11.8 points.

• Ousted Savannah police officer gave ICE agent husband access to Flock system

  A police officer shared her Flock login with her ICE-agent husband; an audit found 983 searches and she was fired

• Tmux and Claude Code: Sessions, Send-Keys, and the Ceiling

  Running Claude Code agents in tmux works well, but limits appear around four agents.

Related Resources

• https://github.com/browser-use/jev-ultrafast