Jev in Practice: Browser Use and Agent Workflows

A developer added a /jev slash command to OpenCode that feeds the last four messages to Jev to answer a yes/no question instantly with a certainty percentage, without using LLM context.

A GLiNER2.5 fine-tune trained as a Jev-shaped computer-use chooser scores 128/128 on the held-out menu test at 36ms on MPS, about 7x faster than hosted Jev, but it only picks from a preset menu and does not plan or click.

LangChain's engineering team is hosting a livestream on how TypeSafe AI's Jev makes fast structured decisions, where it fits in the agent loop, and how to build a harness with it.

The author built jevable.com, a site that gathers Jev demos from X and sorts them into a few filterable categories, with a + button for anyone to submit their own project.

usenaive integrates Jev into a 7-agent agency template covering sales, content, client management and site building, with one config defining the team and routing jobs to different models.

The author notes many people don't understand Jev, offers a three-tier path to learn about it, and sums it up as a fast classifier.

Using old quant infrastructure, the author gave Jev two years of OHLCV plus proprietary data and long/flat/short actions, finding it makes patient decisions, avoids overtrading, and is consistent, cheap and fast.

A developer criticizes the quality of widely shared Jev examples and warns people not to try them at work.

He used Jev to drive a browser for price comparison and found his own pricing page looks good to humans but is structurally poor for bots and screen readers.

He classifies ad topics with a small self-hosted ML classifier in milliseconds at zero cost and publicly questions why Jev would be needed for this.

The author sees Jev's discriminative design as close to the ALIGN work he did three years ago: a large generator paired with a small judge for generalization, efficiency and overall system reliability.

A Turkish developer found diagnosis and fixes are still done by the main Codex or Claude model, with Jev stepping in only at decision points like which hypothesis to test first or whether evidence is sufficient; on a small sample Codex alone scored 20/50 and with Jev 24/50.