Jev in Practice: Browser Use and Agent Workflows

The author will present at Jev Hack/Demo Night, highlighting Interfaze's open-sourced 4B Jev model Lev, which is trending on HuggingFace, with a 27B version and open-source Interfaze to follow.

The author surveys the wave of System 1 decision models sparked by Jev, including OpenAI's Decisions API, AWS Strands Decider 2B and Cloudflare's Clef models, which skip text generation and output decisions and probabilities in a single forward pass.

A trading bot workflow where the LLM develops the thesis, Jev picks from predefined answers with probabilities, code handles risk limits and orders, and ambiguous evidence goes to human review.

Neon built pg_redact, a demo where Jev determines who is viewing a simulated support inbox and a Postgres SQL function redacts PII based on the viewer's role.

browser-use released jev-ultrafast on GitHub, a browser agent implementation driven by a dynamic, indexable action space.

Cloudflare released Clef in 27B and 9B (Clef-flash) sizes under Apache 2.0; it scores candidate options in a single forward pass for tasks like agent tool selection, request routing, and whether to escalate to a human.

A rundown of what Command Code mods can do, including cmd-mod-jev-nudge, which uses Jev to make decisions and can block or rewrite tool calls and prompts.

Awesome Jev Live organizes SDKs, agents, MCP tools, models, and apps by category and evidence level, with its open-source index refreshing every two hours.

A walkthrough of asking three questions about a customer message in the TypeSafe playground — Choice for the team, Score for urgency, Noul for refund intent — then running the thresholds on 50 hand-checked messages.

The author shares an agent tree to paste into Codex, where forks that need no thinking go to Jev for sub-second decisions and large models only handle real disagreements.

Inception's DevRel engineer compiled a list of applications for the Jev decision model, covering coding-agent compaction, routing, guardrails and verification.

In the Pi 1.0 demo, an extension uses Claude Opus for planning, GPT for implementation, and Jev to decide when to switch between them.