Jev in Practice: Browser Use and Agent Workflows

Following Anthropic's docs, the author sets up a team with Opus 5.5 leading, Sonnet 5.5 building and Fable 5.1 attacking, while Jev handles the forks that need no thinking in under half a second so the team only argues over genuinely split judgments.

The author lets Jev read only: 30,000 events a night are scored in parallel by four questions, 97% stop at "read", only 300 write slots are released, at a cost of just over one dollar.

A Chinese creator breaks down his AI pipeline: JEV scores 20 candidate topics each day across seven dimensions and keeps the top ten, while Muse, Codex and others handle their assigned tasks.

The author ran a large-scale comparison on open-ended text drawn from real work scenarios and found Jev matched the other models on quality for consumer text analysis at lower cost and time.

A checklist of nine projects built on Jev, including Opus effort routing, a Stop hook that checks for evidence, skill routing, Codex subagent routing, commit message checks, malicious code scanning, semantic grep and MCP search scoring.

A new example runs yoagent in Rust on Cloudflare Workers, passing every agent tool call through a decision model: Jev at 267ms and $0.042/M versus Clef at 444ms and $0.24/M, with both getting every allow/deny right.

peon is an open-source Pi-based harness that gives models native opcode tools to log in, chat, move, fight and quest in WoW, while Jev decides real-time combat actions like casting, fleeing or kiting.

With Jev, Cloudflare's Clef, open-source Laya and OpenAI's upcoming decisions API all available, the author asks what developers actually use in production.

On day four the developer reported two bugs and tried Jev with Browser Use for automated testing, which is still limited to the DOM and can't test canvas, with the PoC nearly ready to show their lead.

ANT, PAUL, ZERO, CORTEX and FLY split the work overnight, using Jev at $0.042 per million decisions for routine calls, reserving costlier models for hard cases, and ZERO stops itself before deleting a customer.

The author notes that people are busy building with Jev but skip the core step of defining the decision itself, and argues that a bad decision framework makes speed and data volume meaningless.

Using a web crawler animation, the author argues that most agent work is many cheap small judgments, and suggests using Jev for those branches while calling Claude only when truly necessary.