Jev in Practice: Browser Use and Agent Workflows

A roundup of nine open-source projects built on Jev, including jev-router, Foreman, limpet, jev-guard and DiffJury, covering model routing, completion judging, injection defense and PR triage.

The author separates the routing layer: a small decision model ranks actions and routes simple branches in 12ms, escalating only when deeper reasoning is needed; Drex scores 51.73 on the decision index versus 51.67 for Jev 1.13.0 while running 3x faster.

Eikos 27B is quantized to FP8 and run on one RTX Pro 6000 to do the Jev benchmark in a web browser, released open source under MIT.

The author built a small agent that uses Jev to judge each excuse and talk back, at about $0.00003 per judgment, roughly 300x cheaper than a frontier model.

The author feeds each subagent's role to Jev, which picks the execution order, decides in milliseconds, then runs the agents in parallel and passes results forward.

ATG's weekly report says the community integrated the Jev and Meta Muse models into its trading system, and more than five members built and deployed their own trading agents.

An LMSys blog post describes serving scoring models like Open-Jev with SGLang: /v1/score returns scores for the requested labels, shared context is computed once for multiple candidates, and 16-candidate p95 latency drops from 54ms to 20.6ms.

A recap of the first Jev meetup in New York shows Yaven integrating Typesafe into its tool and model routing, bringing cost and latency down to one fifth of the previous level.

Bolna benchmarked Jev on routing for its voice agents using real calls: it agreed with the company's own router on 72.7% of turns, but got stuck about half the time when the conversation needed to move on.

The author built an agent that runs a Jev classifier on every Slack message as it lands and DMs him when his tone turns rude, after first validating it on last week's messages.

The depthfirst team used Jev, a classifier that outputs probabilities, to build a parallel scanning harness that finds nearly as many vulnerabilities on dfbench v1 as GPT-5.6 Luna, escalating only the hardest cases to frontier models.

Jev skips text generation and outputs probabilities for classification and routing tasks; one engineer who replaced ChatGPT Luna with it for classification reported 5 to 18x faster results with greater accuracy.