Jev in Practice: Browser Use and Agent Workflows

The author says they are building a Jev-based routing layer for Hermes to lower their Anthropic API bill, still at the stage of voice-chatting the design through with Claude.

A roundup of four open-source experiments, jev-playground, jev-demo-router, jev-doc-classification and jev-clerk, using Jev for lead routing, ticket assignment, financial PDF classification and bookkeeping agent decisions.

On the same 10-coin strategy, a pure chatbot polling every minute would cost $230 a day, while the author's setup filters with code first, lets Jev screen long or short, and has Opus 5.5 write the thesis only after Jev approves, for $0.93 a day.

A list of 10 open-source projects that use Jev as a decision layer, from browser agents and trading loops to code review and model routing, built on the pattern of LLM generation, agent execution and Jev decisions.

A walkthrough of how Jev decides whether to retry a failed test, route simple tasks to a cheaper path, or escalate to a human at low confidence, so only hard problems reach Opus 5.5.

In a shopping agent making 20 decisions in a row, Jev placed the order in 14.5 seconds with free output, while GPT-6 Sol was still stuck at step 8 when the clip ended.

Clude says Jev has significantly cut its costs, and it is training its own memory model to filter raw noise before passing data to Jev, announcing it will keep building after TypeSafe's funding news.

Jev makes the structured decisions on 100k car listings, LangChain deep agents research the gaps, and simulated users ask their own questions, with every answer and gap stored.

The author shares a single-page configuration that uses Jev as the decision layer and Opus 5.5 for reasoning inside an Obsidian workflow, claiming 10x faster chats and 80% less time spent on work.

A Sonnet 5.5 agent handles email, invoices and reports while Jev blocks high-risk actions such as amounts over $500, contracts and prices for human review, letting the rest through automatically.

Jev outputs only BUY/SELL/HOLD plus a probability, three open-source trading repos already integrate it, and input costs $0.042 per million tokens.

Diogo Amogo's PDF explains how to rebuild an existing coding agent as a Jev harness, with tests showing up to 200x faster and 400x cheaper results, and it can be handed to Claude Code or Codex to follow.