Jev in Practice: Browser Use and Agent Workflows

The author says a PPO policy took a week of work to clear Super Mario Bros World 1-1, while Jev reached the same result in less than 30 minutes.

The author released the jev-browser repo, which needs Codex with browser tools, Node 22+ and a TypeSafe API key, with the install prompt in the README.

A walkthrough for using Jev on Vercel AI Gateway: create a key, install ai@7.0.105+, call experimental_evaluate with state and questions, and get Boolean, Choice and Score results in parallel for tool selection, risk scoring and output checks.

The author built a model router with Typesafe's Jev that decides which model best fits a request and forwards it there, with a demo video.

A Japanese user demonstrates running Typesafe Jev locally, saying it works but still has room for speed improvements.

The author combined Jev with pi's new markdown renderer to build a pi extension that toggles how much detail appears in agent output, motivated by not understanding what Opus 5 was getting at.

An explainer describes Jev as generating no text and returning only structured judgment values, so format breakdown and hallucination can't occur by design, and suggests simulating it with ChatGPT.

Jev says its output tokens are free forever and that it is 74.9x faster than GPT-5.6 on the same 27 questions, aimed at ticket routing, data scoring, LLM output checks and real-time agents.

An open-source adapter built around Codex's browser tools uses Jev for the DOM-to-action decision loop, with postcondition checks and multi-tab support, and requires Jev access.

A joking complaint notes that people are using Jev for tasks that could be handled by a self-hosted custom BERT finetune or by waiting for a cheaper, faster model.

Given the query "I scoop scoops and sprinkle sprinkles," Jev scored it against 1,016 O*NET job options and returned Fast Food and Counter Workers.

The author argues routing to Claude Code, Codex, or Opencode should move out of system prompts and into Jev, which returns only a typed decision with a confidence score, claiming it is 193x faster and 444x cheaper.