Jev in Practice: Browser Use and Agent Workflows

The author argues Jev represents a new kind of model that doesn't chat or write code but takes state and a set of allowed decisions and returns typed answers with probabilities, suited to routing, filtering, scoring and moderation.

Responding to reader requests, the author starts a long thread on the new Jev model, beginning with Diogo Almeida's background at OpenAI working on ChatGPT and RLHF.

With thousands of API endpoints, using Jev to pick the best one instantly at very low cost makes an agent's routing decisions fast and cheap.

After trying Jev, the author lists 10 product directions that replace traditional LLMs with small typed decisions, such as agent completion checks, claim verification and context admission.

A walkthrough of Jev, which only makes judgments and never writes content, letting you bucket hundreds of emails, comments or job posts at once with confidence scores, plus the full steps to sign up, create a key and install the skill.

Taking a 1,000-job sample from a niche job board, using Jev to tag each listing and judge whether it is real, spam, expired or a recruiter post, at a cost of about $0.10, and building a small dashboard to view the results.

A ten-minute setup using the typesafe-sdk, where Choice, Score and Noul evaluate 40 hooks at once so only 3 reach the real model, with an upgrade order included.

A video breakdown of Jev as a decision layer sitting between an agent and every model, tool, browser action and human escalation, covering 40-200x speed, the choice/score/null primitives and why narrow models help.

Give Jev a market snapshot and a specific question and it returns only a decision rather than long text, while code handles pricing, execution and risk; a Jev trading bot already makes decisions every 300ms on Monad/Kuru, with API costs around $10–25 per month.

The author argues Jev separates judgment from language generation, confirming that mature workloads hand deterministic logic back to code, and that first-generation agent architectures are being rapidly surpassed.

The author integrates Jev as a core primitive in a custom harness for classification, control flow, scaled labeling, structured decisions and dynamic context management, unlocking features previously limited by cost or latency.

A developer notes Jev is not autoregressive so it cannot write code or text, currently supports text only, and can only choose from a predefined action set, but it is extremely fast and cheap and suits state-to-action models for browser and computer use.