Jev in Practice: Browser Use and Agent Workflows

The author built a Jev VJ demo where clips and effects are described in advance, Jev recommends the best-matching combination based on what he says, and he switches them manually.

The Mastra team used Jev to build a workflows classifier, tool approval classifier checks, classifier-as-judge scorers and a classifier processor, and is considering a model router.

The copilot listens to sales calls live, suggests what to say next, helps handle objections, tracks the call stage and gives a live probability of closing.

The author put Jev's computer use to the test to see whether the demos circulating on X reflect real capability or are misleading, and found the results surprising.

The official djev-run repo spins up a Jev API-compatible endpoint on Google Cloud Run with one command, at 35-60 ms per step and about 100-123 req/s at batch@32, costing roughly $3/hour and dropping to zero when idle.

The author put Jev in front of the model menu as an experiment, sending cheap questions to the 2¢ tier and hard ones to Opus 5.5, and shows the cost of each call.

Replying to a user, Langfuse says that with Jev, evaluation is less a series of parallel evals and more a single chained decision flow.

In a test of basic browser workflows, more real-world end-to-end tasks and some computer use tasks, Luna succeeded 96.5% of the time overall while the best Jev implementation found succeeded only 37.3%.

The author gathered Jev use cases scattered across X, GitHub, and Reddit into a tracker sortable by category, stars, and views, and is still collecting new examples.

Tired of the rotating model leaderboards, the author argues Jev's structured approach matters more and that models should be auto-routed anyway.

A community plugin uses Jev to judge which tool calls and results to retain; in one experiment it matched the reference pass rate but dropped a key piece of evidence.

The official account is running an event where adopted agents get a free Jev brain and then compete against each other live.