Jev in Practice: Browser Use and Agent Workflows

Command hooks block rm -rf and .env reads, Haiku prompt hooks judge whether a Stop is really done, and an Opus agent hook reviews diffs before push or publish; the post says Jev resolves micro-forks in code in 15ms.

Omnirail introduces Jev decision models that answer typed yes/no, choice or score questions with a probability like 0.98 go or 0.01 stop, at a fraction of a cent per check.

A roundup of Jev-powered builds: Browser Use's 7-second flight-booking agent, Jev-scored context compaction, jevgrep code search at ~30% lower cost, a Monad market-making bot deciding each block, and pg-jev natural-language WHERE clauses.

A Claude Code bugfix workflow where Haiku triages, Sonnet reproduces, and Opus patches, with Jev engineering resolving 2,340 runtime forks in code at 0.98 sharpness.

Six AI tools including Opus, Grok, and Jev are chained into a team that finds a need, builds a product, and promotes it unattended for 12 hours, claiming 47 paying customers.

The author lists Jev's claimed 70-500ms decisions and 40-400x cost savings, then demands eval tables, latency histograms, decision logs and backtest comparisons.

Developers open-sourced HarnessRouter, letting Grok Bot delegate tasks to Codex, Claude Code, System One (Jev) and other harnesses, demoed by turning a support-ticket CSV into a dashboard.

Routing the 91% of agent calls that are yes/no, source-pick, or score decisions to Jev for a probability and leaving writing to Kimi K3 cuts decision cost roughly 250x.

Swapping only the model behind Claude Code to Kimi K3.1 Pro yields a real 43% saving, with JEV resolving micro-forks in code before escalation.

A researcher discusses decision models for lightweight formal methods and SRE, noting students have tried Jev on the SREGym benchmark and written up their experience in blog posts.

The author just finished a voice mode for their demoed Jev-integrated computer use system and plans to bring that voice layer into poly.

DeepEval adds JevEval, replacing generative LLM judges with Jev's structured decisions across yes/no, graded Score and Choice, including agent trace evaluation.