Jev in Practice: Browser Use and Agent Workflows

A 16-minute live coding demo wires Jev into Mastra for workflow branches, guardrails, tool pre-selection and classifier judgments, with the author's own tests ranking Jev third among the tool retrieval options he tried.

Browser Use built jev-ultrafast on Jev, turning clickable page elements into a numbered table so Jev decides the action and element in one request, with no screenshots and a small model called only for typing; a Google Flights Zurich-to-London search dropped from 9.45s to 7.09s median, and browser communications from 1092 to 101.

An options system has GPT extract claims from earnings materials, Jev classify the evidence type, and QuantLib reprice the trade, so a candidate that gets direction right but is crushed by a volatility collapse is blocked before execution.

The piece explains that Jev is a general decision model for the many small judgments inside agents, and that its calibrated probabilities support tiered handling from automatic execution to human review.

A searchable map of 1,248 studies that use 27 kinds of Japanese public statistics, such as vital statistics, NDB open data and suicide statistics, browsable by theme, data source and data year, built with TypeSafe AI's Jev.

OODA AI runs a full Computer Use Google Flight search in OpenClaw in 7.1 seconds using Sol as primary model, Luna as light model and Jev as classifier, and says this cuts context use by 50%.

A CTO ran 34 handwritten and 1,934 real production requests on his home desktop and concluded Jev's speed comes from single-token decoding rather than the model itself, and that with zero labels it ties a plain 27B open-weight model on accuracy.

The author shares how a Vercel recruiting DM reached him: the recruiter saw his viral Jev demo, read his posts about agents, then followed his personal site to find him.

The author open-sourced a demo repo that uses Jev for browser use, runnable directly to see how the decision model drives browser actions.

It combines a harness inner loop with a rubric-guided RL outer loop, offering free GPT-6, GPU acceleration and the JEV framework to scientific agents.

He describes Jev as a classifier rather than a chat LLM: it makes decisions in about 200ms at very low cost, making browser-use tasks much faster and cheaper than letting a model click around.

The author uses Sai to pretrain moves and Jev to pick the best move in real time, letting an agent play The King of Fighters on Windows by itself; it lost the first match and won the second.