Jev in Practice: Browser Use and Agent Workflows

The author recalls Roast, built by Shopify Engineering last year, and says it is a perfect fit for Jev, linking to the official introduction post.

A prompt to give Claude or Codex builds an agent that uses Jev to pick models by query difficulty, with the interface showing each routing choice as it happens.

A single Decision request scores refund probability, picks a department from a closed set, and grades frustration, while PHP still controls the action policy.

A toy project by _rchaves_ restricts Jev to caveman words and, like an LLM, forces it to pick the next word one at a time.

A repost says Browser Use combined with Jev builds a new action space at every step and found flights in 7 seconds at a cost of $0.0039.

A developer made a first-person shooter where the player duels Jev one-on-one, and it is live and playable.

Isaac Flath lists six Jev uses he expects to still rely on in 60 days: fact-checking scripts, ranking news, finding text in PDFs, checking citations, grouping review notes, and evaluating agent failures over traces.

A Netlify demo hands program state to Jev and accepts only preset answers like sales, support, or spam, so it can branch directly without parsing text.

The author shows a real Jev request for a Browser Use scenario, with state holding the page and recent actions and questions asking for the operation, click target, input value, and whether the task is complete.

A Monad developer built an on-chain trading bot where Jev makes pure classification decisions, open-sourced the code and frontend, and the demo lost over 300% in 14 hours.

The author catalogued 46 public projects by what decision Jev makes, including Notra's production classifier switch, a browser-use Jev fork, and Vercel eve's default eval model.

The author shows how Jev can inject skills and system prompt guidance just in time, such as nudging an agent that is too verbose or drifting from AGENT.md rules.