Jev in Practice: Browser Use and Agent Workflows

A trading bot lets Grok guess, QuantGPT test, Jev judge and code place orders; a rule that scored 0.86 in-sample but 0.17 out-of-sample was marked NO EDGE and discarded.

A Claude Code setup runs Sonnet on the main session, Haiku for parallel exploration and Opus 5.5 as advisor, while Jev handles deterministic forks like tool choice, paths and retries at millisecond speed.

The author runs agents as desks with a switchboard and a small-model front desk, using Jev classifiers for routing, digests, notification filtering and shared context across desks.

A list of eight tools to cut Claude Code bills, including jev-ultrafast, which lets Jev decide each click and only wakes a text model when typing is needed.

600 Claude agents in 30 squads ran five task types, with Jev scoring each output and holding anything below 0.90 for human review, so four passed automatically and one was held.

Short multiple-choice decisions like which notes to load, whether to retry, and which tests to run first go to the cheaper Jev model, while Opus 5.5 keeps planning, coding, and debugging.

Three Claude Code loop tasks maintain the open-source project, and one uses Jev to first judge whether a tweet is relevant so unrelated content isn't sent to the expensive model.

The Praxis team applies Jev to law firm inboxes, classifying whether a message is a new lead or existing client, which matter it belongs to, and whether it is a document or payment, with a confidence score.

The author recommends clef-flash for the 8–16GB tier, saying it works like Jev and suits browser use, automation and high-volume processing.

A breakdown of how Jev only returns choices, scores or probabilities, gates risk before tool calls, and cut agent context by 91.5% in a compaction test.

A Claude Code kit that uses Jev to check whether a subtask could go to a cheaper agent, blocking and rerouting when it is too expensive, and can also drive browser clicks.

Wiring Jev into Claude Code for classification, scoring and checks so simple yes/no calls stop burning Claude tokens, which the author says cut costs about 30% on the same tasks.