Jev in Practice: Browser Use and Agent Workflows

A toolbox of 20 GitHub repos covering TypeSafe skills, routing, caching, and cost tracking to replace routine Opus decisions with Jev.

Anthropic engineer Boris Cherny says 90% of engineers run agents with self-improving loops and recommends a loop engineering guide combining Opus 5.5 with Jev.

Structured questions like lead follow-ups, email replies, document reading, specialist routing, and approve/reject/escalate go to Jev, while Claude still handles research and writing.

Six seats including STRAT, EXEC, and SCOUT use Jev as a yes/no judge; the paper run made a daily profit but still lost to simply holding, with a transparent, reviewable process.

In a small test case, Jev routed GPT 5.1 Medium tasks to open models on Nebius, showing enterprise-level cost savings.

Compares the cost of 1,000 yes/no decisions across Astra, Opus and Jev, and suggests using Jev for review, routing, lead classification and quality scoring.

Jev decides who to assign, whether a lead is worth following up, whether an email needs a reply, whether a source is worth reading, and whether to approve, reject or escalate, while Claude does the research and coding.

Jev does the first pass on spam, duplicates, whether a reply is worthwhile and priority, and Claude only steps in when something actually needs to be written.

Install jev-mcp and jevgrep so Claude Code hands easy questions to Jev for fast judgment and only wakes Claude for the hard ones.

After the amount is locked, Jev selects the service, the registry supplies the payee, and Kimi writes the reply; live on Monad mainnet.

Seat 1 is Jev, which asks whether flow is toxic, what the regime is, and who should think; the annual bill drops from $12,614 paid to Opus for saying no to about $394.

Move choice, score, and yes/no questions off the LLM; the author lays out an order from auditing calls to the three question types.