Jev in Practice: Browser Use and Agent Workflows

The author discusses TypeSafe's Jev model in an X long-form article, with the specific claims available by opening the post.

Using Jev to map each step's DOM state into an action space, the agent found flights in 7 seconds at $0.0039, with a small LLM only as fallback for typing.

The author rebuilt a job crawler that finds a company's Careers page and scores matching roles, and the Jev version ran about 15x faster than the LLM version; it is packaged as the jev-job-hunter skill.

Following the official steps, the author joined the waitlist, installed the agent skill and got an API key, then used Jev on pollution and debugging tasks in their own app, calling trading bots a typical use case.

The author moved an AI review feature from GPT models to Jev and got the same results in seconds instead of about two minutes, describing Jev as a classifier that works out of the box from a prompt with no training.

A long roundup of real Jev uses and limits: a booking agent finds flights in 7 seconds, a trading setup decides every 300ms and a Doom demo acts every 90ms, but none of them see images, relying on text state built beforehand, and speed multipliers vary by test while email classification lost to a Google model.

A developer who planned to argue against using LLMs for classification finds Jev returns typed decisions instead of text, and weighs its speed and cost against the taxonomy work it leaves to the user.

A developer built JevScope, which asks Jev at each step whether an agent's action is aligned with the task, making progress, or repeating itself, and plots the result over time.

After testing Jev, a user argues it suits the position-monitoring judgments a numeric risk system can't make, and imagines Noesis agents asking its API whether a meme coin should still be held.

The author built an agent router tool using Jev and herdr, recorded a YouTube video about it, and crossed 10k views in under 19 hours, their best result on the channel so far.

A linked list of Jev demos covering viral post scoring, an on-chain trading bot on Monad/Kuru at 300ms, Browser Use with Stagehand, a Google Flights query in about 7.1 seconds, and agent memory compression.

Someone shared a screenshot of Okta adopting Jev, but the post gives no specifics, so the use case has to be read from the image.