Jev in Practice: Browser Use and Agent Workflows

The author reviews six papers and tests showing how zero-shot judgment models like Jev can handle classification, memory selection, action choice and first-pass answer review for agents, saving many small decision costs.

RWKV officially lists 23 agent projects, with Jev named among them, and the directory shows what each project does.

A roundup of six open-source projects using Jev as their decision layer: browser automation, a Monad trading bot, a Claude Code/Codex safety harness, an MCP bridge, a tax form classifier and a Postgres extension.

5,400 agent runs completed unattended: 43 Jev decisions across 18 turns, a $0.02 cost per decision and a goal score of 0.93.

A demo of the state→decide→dispatch→fleet→score loop: 512 agents run in parallel at 0.4 seconds per step, with no human approvals and the run stopping itself after 18 turns.

In comparison tests, GPT-6 Astra was persuaded by a single sentence, while Jev never reads text and only scores it, flagging instructions hidden in a $4 million transfer request as fraud in 40 milliseconds.

A CC0 open-source GitHub directory that organizes Jev SDKs, docs, agent workflows, and browser-use examples by what you want to build, helping developers find a starting path quickly.

The author built a site that redesigns any pasted link: Jev returns a structured decision (layout, copy, palette) in about 0.1 seconds, and Opus 5.5 writes the TSX component for each section as you scroll to it.

Across 240 paired runs on browser-use, WebMCP, and NLWeb, adding the Jev decision model kept task success unchanged while making agents 3.6x faster and 7.7x cheaper on average.

A developer built a smart home controller with Jev inside GitHub Copilot and called it simple, fast and cheap, saying the hype is well deserved.

Diogo Almeida explains how to build evals around your own workflow, use cascades for cheap decisions, and build coding agents with Jev, with timestamped sections.

A Superlinear podcast episode discusses placing Jev between deterministic code and expensive LLM reasoning, covering browser loops, real-time feedback, test selection and model routing experiments.