Jev in Practice: Browser Use and Agent Workflows

The author used Jev to score 3,000 hotels and flights in parallel, filtering fake reviews and hidden fees, then had Opus 5.5 plan only the top 5, for a total cost of $0.25.

In a computer-use test, Jev drove the screen with a fast model for routine steps and handed a step back to Opus when its confidence dropped, forming a fast-slow model relay.

You.com used Jev as a reranking layer between retrieval and synthesis in the Knowledge feature of its Web Search API, cutting token usage to one third and reaching 84% accuracy on the Vertical RTK benchmark.

In a launch-week breakdown, Rob Shocks reported Jev at about $42 per billion tokens and roughly 100ms per answer, with a smart home demo returning an executable choice in 185ms.

The author argues that LangGraph's core abstraction of putting decisions into state and routing on them fits Jev well, and may come to be seen as the general way to build software with composable models.

Paddock 0.1.10 supports Windows, Linux, DGX Spark and macOS, and adds a Jev-style classification API built on DiffusionGemma 26B-A4B, along with Laya and Qwen Image-2.1 models.

hwchase17 says modeling agent systems with LangGraph and embedding Jev to speed up the many small decisions is a great combination.

Abnormal held a 100-person R&D offsite in Houston, where the author's recap mentions building a hackathon project with Jev alongside colleagues, plus roadmap sharing and team bonding.

Moving decisions like picking knowledge base entries, guessing user intent and choosing retrieval methods from the LLM to Jev brought a docs agent's median response down to 0.5 seconds.

A circulated 14-block architecture diagram assigns planning, research, coding, execution and review to separate agents, with Jev holding decision memory across them; it claims an 87% task closure rate versus 20% for simple prompting.

A coding agent architecture pairs Jev with Claude Code so cheap judgments handle gates and routing, escalating only high-risk or uncertain decisions to Claude or Codex before re-running verification.

Jev is presented as an engine returning typed decisions like BLOCK, ROUTE, SCORE and VERIFY, with steps to run the browser demo, call a /decide endpoint, and build browser agent and trading loops.