Jev in Practice: Browser Use and Agent Workflows

The author compares SAM to Jev for segmentation and imagines locally runnable open-weight specialists for web navigation, data viz and video editing, orchestrated by a generalist agent.

In a voice-controlled home setup, Jev judges after every word and requires two approvals in a row before acting, so devices change state one by one as they are named.

The author says he tested major classifiers including Jev and Cloudflare Clef and argues they are overhyped, with fundamental flaws when used to guide key pipeline decisions.

Beacon records sessions from Claude Code, Codex, Cursor and 20+ other tools, while Jev extracts lessons worth remembering so one agent's mistake can be learned by another.

Deterministic PreToolUse, PostToolUse and Stop hooks in .claude/settings.json block unsafe commands and enforce tests, while Jev applies the same idea one layer down by resolving no-thinker forks in code in milliseconds.

New research turns user activity into a knowledge graph where agents propose candidates for decision models like Jev to select from, an approach the author says agent builders should explore.

The author trained TERN-35B-A3B to explore the Jev decision model's price frontier, reaching $0.0116/Mtok on an H100 PCIE and $0.00773/Mtok after optimizations.

The author improved a workflow DSL that turns agentic loops into deterministic recipes, leaning on TypeSafe's Jev for routing, classification and composition to cut agent cost without losing performance.

A Grok bot manages the ATS and pre-screens candidates, then sends seven questions whose answers Jev evaluates to shortlist for real interviews; about 37.5% respond and only 4% pass, with the full workflow and funnel report shown on video.

Explains how Claude Code SKILL.md files keep only descriptions in context and load the body on match, while Jev routes the forks that need no thinker in milliseconds.

The post lists early Jev uses: 500 emails sorted for 3.5 cents, a browser agent finding flights in 7 seconds, a trading bot deciding every 81ms, and a Claude session compacted from ~1M to 86K tokens.

A Tsinghua student wired OpenAI, Jev and Grok Bots into one Slack channel plus gate.py to follow up 40 leads, with Jev routing and answering yes/no questions and high-risk items escalated to a human.