Jev in Practice: Browser Use and Agent Workflows

The post points to the official Jev repository, saying it covers model routing, guardrails, multi-agent workflows, context compaction, tools and skills.

The author tested merging 10 questions into a single Jev call, cutting latency from 6.6 seconds to 0.66 seconds and input tokens from 62k to 6.2k, and gives the rule: same state with no dependencies means the same call.

Code handles deterministic rules, Jev uses a $0.04 model to decide which tasks are worth escalating, and Claude handles only the remaining ~10% of judgment calls, bringing the daily bill from $180 to $19.

The author gave Jev Bot $55 and told it to pay for itself; in 20 hours it grew to $8,960 by scanning markets every 9 minutes, computing thresholds, sizing bets with the Kelly criterion and covering its own hosting cost.

A six-step workflow where Astra turns meeting notes into markdown pages while Jev checks for duplicates, flags conflicts and picks files, lifting retrieval accuracy from 40% to 96%.

The roughly 1,300-word skill has Claude Code read live docs first, keep rules in normal code and use Jev only for judgment, with a test showing 13 questions in one call cost 12.2x less.

The author shows Jev returning both an answer and a confidence score per call, with thresholds set by risk: 0.5 for read-only, 0.7 for refunds, 0.8 for escalation and 0.9 for destructive actions, with anything below 0.5 going to a human.

A roundup of nine open-source builds using Jev for decisions, from smart-home queries and filter-list-free ad blocking to a git commit gate, Pong and a Hyperliquid trading bot.

Jev routes, scores urgency, checks evidence and safety for each event so only high-value content reaches Grok Bot, with a cost test at 20,000 decisions per day.

After testing Jev, he says it is more than a classifier and can hold complex state, but its accuracy does not yet beat frontier LLMs, and he predicts everyone will use this kind of model next year.

The official playbook replaces prompt writing with question design: one judgment per question, field paths, contrastive criteria, escape hatches, speculative fan-out, confidence routing and per-dimension scoring, plus 5 hard rules.

The author shares a single prompt that installs Jev, audits Opus's context and token waste, and moves routing, scoring and verification to Jev while Opus keeps the deep reasoning.