Jev in Practice: Browser Use and Agent Workflows

AgentRun is a new harness built with Jev that lets an agent learn a job, code itself a general solution, and then step out of the way.

The author built a Claude and Pi plugin that uses a Jev classifier to judge whether the agent is at a task boundary safe to compact, with hint and auto modes, evaluated on 40 real sessions.

After trying Jev, the author says the most useful use cases so far are generative UI and computer use or browser agents for QA.

Cua built a 2.8MB, 706K-parameter form-filling model that scored 99.7% on its eval, versus 83.6% for hosted Jev, and runs locally.

fkadev argues that big labs could ship a comparable structured-decision model inside APIs developers already use, and that distribution advantage could make it hard for TypeSafe/Jev to compete even with a strong product.

Someone shares an idea for using Jev to generate game levels in real time, showing how it can drive dynamic content generation.

The author argues that once Jev makes generative UI cheap and fast, plugins will ship with UI components and company interfaces will shift from dashboards to components assembled on the fly.

A short demo video uses Jev to make a simple yes/no decision, showing how it handles structured binary choices.

A team built an incident triage workflow that feeds unstructured content into Jev and gets typed decisions back.

A developer made a browser FPS demo where the Jev and OpenJev models fight each other, playable directly in the browser.

Jared Palmer trained a tiny decision model on Qwen2.5-0.5B with a TypeSafe-compatible API, and released the weights and model card on GitHub.

A small experiment compares two versions of a document to show whether an edit changed the meaning or only the wording, returning results in seconds.