Jev in Practice: Browser Use and Agent Workflows

SERV v3's Graph Sharding breaks reasoning graphs into script, Jev decision, and constrained-LLM nodes with logged cost per step; the team reports SERV plus a small model beating a frontier model at about 90x lower cost.

A trader routes every new pump.fun coin through a Jev gate, read-only OpenAI wallet checks, an Opus 5.5 verdict, and a final human tap before buying.

The platform now returns probabilities for agent questions; in the demo rm -rf ./dist scores 0.60 to run while rm -rf ~/ scores 0.01 to block.

Lightpanda embeds Jev in its binary as judge(), a ~0.2s call that tells whether a page has real content and what kind, avoiding sending whole pages to an LLM just to spot cookie walls.

A Jev 1.13 scanner decides whether 84 credentials of departed staff are still needed, while Sonnet 5.5 reads logs, Opus 5.5 drafts revokes and a GPT-6 Dot pings managers, all in 40 minutes for $6.38.

The author hands multiple-choice decisions to Jev while Opus keeps hard reasoning, with each pick returning in ~100ms and free output tokens, saving about $3,180 a month.

An open-source Browser Use agent has the model pick actions and elements from a numbered list instead of writing selectors, demoing a 7-second flight search as a research prototype.

Trading tool queen harness shipped a built-in chain indexer that uses Jev to label wallets as bundle wallets, bots, or smart money/KOLs, plus holder clustering and dev history.

An Apache-2.0 27B decision model that returns a probability per option for agent routing, tool selection and out-of-scope detection, runnable locally with vLLM.

Compares Jev, Decider, Clef and other decision models on 98 production agent traces, showing threshold tuning, calibration and cost-performance tradeoffs yield free, measurable eval gains.

An open-source crawler checks 52 SEO rules, then asks Jev to classify each page's type and search intent, producing evidence-backed prioritized reports.

Hamel tests Jev and other decision models on 98 real agent traces, showing threshold tuning shifts false positives and negatives, and advises calibrating on labeled data first.