Jev in Practice: Browser Use and Agent Workflows

Using a duplicate-charge complaint as an example, the author argues that instead of having a chat model emit JSON to parse and retry, you define the possible answers up front and Jev returns structured decisions and probabilities your code can use directly.

The team shows a trading bot built on Jev that feeds in live data and lets agents place real trades within the user's risk limits.

The post explains that Jev returns only decisions with probabilities for each option and is billed as a System One model, and that a Thai team built OpenThai-SystemOne, a 0.8B Thai-language equivalent, within five days.

The post traces Jev's name beyond the Jevons paradox to the logic machine Jevons built, which did decisions rather than arithmetic, matching Jev's positioning.

A video demos three apps built with Jev: sorting 100 comments, grading a blog draft as you type, and picking a model before any tokens stream.

On the same 600 paragraphs and 176 features, Jev answered nearly all of them but got 193 wrong, while Slop Meter answered one in nine and got a single one wrong, which the author uses to argue for saying "I can't tell."

The open-source long-running agent control plane LoopX has an RFC that hands goals and code changes to Jev to judge whether work is drifting from the goal, off by default and able to run in observe-only mode.

Jev adjusts reasoning depth at millisecond scale along the execution path, thinking less when things flow and more when stuck, with tests showing about 50% lower cost and no loss of existing cache.

A commenter argues Jev is still a hosted API that enterprises would only adopt with a self-hosted version, and that deterministic decision engines have long existed as internal components, so handing simple decisions to LLMs is itself a design flaw.

The author open-sourced a browser skill: given a starting URL and a goal, Jev drives an isolated Chromium instance, clicking and navigating to the target page, and it works with Codex and Claude Code.

After running real evals, the author finds Jev performs well on browser and computer use and is even better as an LLM-as-a-Judge replacement, with a hybrid agent where the LLM does the work and Jev makes decisions as the next step.

After reviewing a Jev paper-trading demo, the author concludes it cannot anticipate trends and would lose money trading independently, so Noesis uses Jev mainly for signal discovery and risk control, triggering investments directly only on exceptionally strong signals.