Jev in Practice: Browser Use and Agent Workflows

A 27K-parameter model turns natural-language queries into structured filters in the browser in 0.25 ms, handing only uncertain words to Jev.

The author describes fast-jev-compaction, which uses Jev to quickly decide which tool calls and results in an AI coding history can be dropped rather than summarizing, keeping key constraints and error evidence verbatim and cutting 156k tokens to 62k in a demo.

Marcel Pociot built a macOS app that watches the Downloads folder and uses Jev, with no other LLM calls, to decide whether a file is an invoice and rename and file it accordingly.

A Chinese tutorial walks through four steps: join the waitlist and fill out the questionnaire, have an agent install the skill, create an API key, then run the TypeSafe skill, and reports very fast responses.

The author says System One-style models beat frontier models at browser and computer use, are extremely fast and nearly free, and will greatly change how people interact with machines.

The experiment feeds Jev price data, funding rates, open interest, liquidity and macro data, and it picks an action such as buy, sell or hold with a confidence score, which code then executes.

The Jev ecosystem list grew from 61 to 100 entries, adding projects such as Foreman for acceptance checks, a 0.6B NanoJev replica, and typesafe-adblock.

The suggestion is to repeat the same flight search, change dates and filters, and rerun after a page layout change to see whether 7 seconds at $0.0039 is actually useful.

fatwang2 built Jev Search, an open-source search front end and tool on Typesafe Jev that uses natural language to choose retrieval sources and order results, available at jev.s1.dev.

Jev evaluates predefined options in parallel and returns typed values with probability distributions; in the demo a browser agent uses it to find a Zürich to London flight in 7.1 seconds.

Jev is now available on Aria, either through a Model API key inside the platform or via a direct decision API call from your own agent, for choosing models, scoring options, routing tasks, detecting risky actions and deciding when a human should confirm.

zanifold used Jev inside Minis to drive decisions for minis-browser-use, reaching 100% accuracy across 21 test cases in 16 seconds at a total cost of $0.0019.