Jev in Practice: Browser Use and Agent Workflows

A Jane Street quant reportedly shared a "Jev Judgement Battery" cheatsheet on Reddit for building a 24/7 HFT trading system, claiming Jev makes accurate decisions in under 100ms.

A satirical meme post claims that after using Jev, code quality rose 117% and the author can see through walls and the 5G spectrum, meant as community banter rather than a real review.

The author argues Jev's typed choices and scores are more reliable than asking an LLM for JSON, but notes that "zero hallucination" only guarantees format, not that the decision is correct.

Ryan Vogel demos Jev routing 1,700 emails by confidence, with 4.2M input tokens costing just 18 cents, and shares a 13-question request pattern that cut cost another 12x.

The author open-sourced Jeff-1, a small local model imitating how Jev makes decisions, with a training guide, and says it is overconfident and best used as a starting point for improvement.

A single evaluate() call runs 18 questions in parallel, classifying objections and competitor signals, and high-confidence results automatically trigger scripts or pull in human support.

The author built a real-time trading bot with Jev in an evening and morning that ingests onchain and offchain data to make autonomous trade decisions, though it has lost $31,680 so far.

The author predicts Jev plus CUA will produce many specialist models each doing one narrow job, such as scoring, retry decisions, or form filling, with a frontier model sitting on top.

The author curates 20 of the 287 Jev open-source projects listed on their directory site, grouped into ones that run right away, ones that show Jev's advantages clearly, and ones with wild ideas, covering browser agents, context compression, real-time UI, trading bots, Mario and drones.

The author embedded Jev into a general browser skill so that, given a URL and a goal, Jev handles the decisions along the way, and asks what others would automate with it.

The team packaged Jev into a browser skill called jev-browser: give it a website and a task, the browser opens automatically, and Jev decides every click based on what is on screen, with a demo video attached.

Each conversation is split into a checkpoint graph at sentence and moment boundaries, an agent harness pushes artifacts to trigger each Jev classifier, and a custom renderer displays the results.