Jev in Practice: Browser Use and Agent Workflows

The author says pairing Jev with Hermes covers routing, memory, skill selection and GUI clicks at about 0.4 seconds and under a cent per call, and has open-sourced the repo.

The author lists 21 cases where Jev replaces large models for small decisions, including model routing, tool selection, retry logic, a spend firewall and human review queues, with a build approach for each.

The author applies Jev to outbound sales, combining Saleshandy company data to judge which accounts likely have the problem being solved, explain the reasoning and rank the top 100.

An incoming prompt is mapped to an eval, then Jev picks the model with the best cost and completion time for that eval, either deciding automatically or showing the cost-latency tradeoff to the user.

The author released a demo repo that uses Effect's Jev support to classify vulnerability descriptions against CWE IDs, with runnable example code.

With type=choice and fixed option descriptions, tweaking how State is written changes which option Jev picks and how confident it is.

A post asks what the coolest thing built with Jev is so far, with the author naming real-time clipping and browser control as the first use case that made things click.

A video covers someone who automated a trading bot with Jev and lost $32k, showing a real-world failure case.

On the same NetSuite saved search task, Trouve Browser Agent with Jev finished in about 45 seconds while Codex Computer Use took 11 minutes 30 seconds, one user says.

He argues Jev is over 400x cheaper than general LLMs, making ad-supported AI-native apps economically viable for the first time, and asks for examples of new apps built on Jev.

A prompt asks ChatGPT connected to GitHub to inventory LLM call sites across all repositories, calculate potential savings and ROI from switching to Jev, and open GitHub issues directly.

The author replies that Jev still needs an agentic loop layered on top to work, but believes Jev paired with anything will be powerful.