Jev in Practice: Browser Use and Agent Workflows

The agent reads a DOM table and picks click, type, or select actions, calling a small LLM only when it needs to type, and booked the flight for $0.0039.

A reposted thread claims Jev makes agents 10x faster and cheaper by routing each task to the right model without retraining.

Responding to thdxr, a developer says Jev should live in the harness handling routing and scoring for free, which is what it was built for.

tobi shared a demo that runs Jev directly in the browser, so the decision model can be tried without a backend.

The author uses Jev to generate a dataset for training a mini-Jev on Qwen 3.5 2B, testing how well RLCD works for small language models.

Inspired by Jev, a Shopify engineer built Reflex, a Qwen-based experiment in structured decisions with probability outputs that runs on WebGPU and is now live.

The poster says people will soon use typesafeai's Jev to expand queries and rerank results, making RAG work well again.

A long post summarizes Archer Hume's black-box testing of Jev: probabilities are read directly rather than generated token by token, the backbone is a causal Transformer, calibration error is 0.0313, it is likely a sparse MoE, and it explains which agent decision points it suits.

The author questions the moat of Instinct's fine-tuned Kimi model for browser use, argues Jev could change the computer-use landscape, and offers a strategy for application-layer startups.

A roundup of what people built with Jev within a day of release, including Super Mario, Tetris, and a Tesla FSD recreation built in under an hour, showing the range of a model that only makes judgments.

A developer lists the parallel Jev calls Puffle.ai plans to run every turn: whether a turn is worth remembering, whether the task is done, whether an error occurred, whether to route to a small model, and whether to reply on a public channel.

A user ran Jev over a YouTuber's last 100 videos to assign each a sales intent score, finishing in 12 seconds at about $0.02, with accuracy still to be checked.