Jev in Practice: Browser Use and Agent Workflows

The developer replaced the LLM routing layer of the VibeWrite agent with Jev, using structured selection to decide which model handles each request.

The article argues Jev replaces LLM-generated fake confidence with probabilities read directly from internal representations and trained on real outcomes, but its weights are closed and its research unpublished.

The author tests and ranks five open-source decision models that can replace Jev, covering speed, accuracy and hardware needs, and names his usual System One pick.

The author adapted a browser tennis game so two decision models each pick tactics, tracking win rate, response latency and API cost, with Decision-Machine-1 currently ahead.

The open-source project uses Jev to pick operations and elements from fixed actions like CLICK, TYPE_TEXT and SELECT, calling an LLM only when text needs to be written, and completes a Google Flights search demo in about 7.1 seconds.

The open-source project pre-fills the answer structure and leaves only decision slots open, reading DiffusionGemma 26B's probabilities instead of generating JSON token by token, and runs locally to expose a Jev-compatible endpoint.

A video demo shows Jev choosing driving paths in real time in a Tesla Full Self-Driving scenario, with a browser demo link to try it.

The author used 8 years of listening history and 16,558 songs to build recommendations with adaptive reasoning in five steps using Jev and Claude, at a cost of $0.31, with the code open-sourced.

An npx-installable skill that pairs a frontier model as orchestrator with a lighter model for subagents, adding a Jev-style classifier to speed up orchestration decisions and lower token costs.

A 12-step tutorial where Opus generates strategies and Jev turns market state into typed, probability-weighted choices that output ENTER/EXIT/HOLD/NO TRADE.

The poster wants to simulate day trading with Jev to see how quickly an account goes to zero, and is only looking for people to try it, with no results yet.

The author places Jev on the edges between nodes, with each link a typed Jev question, achieving 0 missed transitions and 0 false alarms at 7.7 versus 11.6 lab-hours in a superconductivity cooldown experiment.